Source-linked AI summary
A General Algorithm for Deciding Transportability of Experimental Results
Elias Bareinboim, Judea Pearl
TL;DR
The paper addresses how to generalize experimental causal findings to a different population where only observational data can be collected. It formalizes population differences with selection diagrams and develops graphical criteria and a complete algorithm for deciding transportability and computing the appropriate transport formula. The results provide necessary-and-sufficient decision criteria and formal guarantees, while requiring sufficient background knowledge about how populations may differ.
Problem
Existing work discussed generalizing findings across populations but lacked a formal treatment of when experimental causal knowledge can be transferred to a target population.
Method
The paper represents cross-population differences with selection diagrams and uses causal-graphical analysis to derive decision criteria and an algorithm for computing transport formulas.
Results
The article provides a complete necessary-and-sufficient graphical condition for transportability and an algorithm that computes the correct transport formula whenever the condition is satisfied.
Takeaways & Limitations
The formal framework specifies what assumptions and measurements are needed to infer target-population causal relations from experimental and observational information.
Takeaways & Limitations
The analysis assumes the analyst has sufficient background knowledge to determine, at least qualitatively, where the two populations may differ.
Abstract
from arXiv · showhide
Generalizing empirical findings to new environments, settings, or populations is essential in most scientific explorations. This article treats a particular problem of generalizability, called "transportability", defined as a license to transfer information learned in experimental studies to a different population, on which only observational studies can be conducted. Given a set of assumptions concerning commonalities and differences between the two populations, Pearl and Bareinboim (2011) derived sufficient conditions that permit such transfer to take place. This article summarizes their findings and supplements them with an effective procedure for deciding when and how transportability is feasible. It establishes a necessary and sufficient condition for deciding when causal effects in the target population are estimable from both the statistical information available and the causal information transferred from the experiments. The article further provides a complete algorithm for computing the transport formula, that is, a way of combining observational and experimental information to synthesize bias-free estimate of the desired causal relation. Finally, the article examines the differences between transportability and other variants of generalizability.
1 Introduction
Transportability concerns transferring causal knowledge from an experimental population to a target population where only observational data are available. The article formalizes the assumptions needed to decide and perform this transfer.
- Transporting findings across populations is a pervasive scientific practice, including applying laboratory conclusions to different environments or groups.
- Existing discussions across statistics, economics, and health sciences offered heuristic guidance but lacked a formal treatment of generalizing across populations.
- Using selection diagrams, causal diagrams, and do-calculus, the article develops an effective procedure for deciding when and how transportability is feasible.
- The framework distinguishes experimental causal knowledge in a source population from passive observational information in a potentially different target population.
- Selection diagrams represent population differences through unobserved factors that can generate discrepancies between domains.
- The article presents formal definitions, graphical criteria, and an effective procedure for deciding transportability and deriving transport formulas.
2 Motivation
Three examples show that transport formulas depend on how population differences are embedded in the causal structure, not only on observed distributions. The article then uses do-calculus and selection diagrams to characterize transportability and motivate a complete algorithm.
- Examples: Three examples motivate transportability through differences in age distributions, proxy language skills, and an exposure-dependent biomarker between treatment and outcome.
- Example 1: When age-specific effects are invariant, the target effect combines experimental age-specific effects with the target age distribution.
- Example 2: A proxy for age need not preserve invariant subgroup effects, because age rather than language skills shapes treatment responses across populations.
- Example 3: For an exposure-dependent biomarker, the transport formula weights experimental z-specific effects by the target conditional distribution P*(z|x), rather than by the marginal P*(z).
- Formalization: Lemma 1 characterizes transportability through reducing a selection-conditioned causal expression, with source-estimable terms lacking selection variables and target-estimable terms lacking interventions.
- Algorithmic gap: Lemma 1 verifies a supplied do-calculus reduction but does not determine whether a reduction exists or provide the rule sequence computationally.
- Contributions: The article addresses this gap with a general graphical condition and a complete algorithm that decides transportability and returns a transport formula when feasible.
3 Preliminaries
The preliminaries define structural causal and selection diagrams, identifiability, and transportability for comparing source experiments with target observations. They also identify graphical structures that obstruct transportability.
- Structural causal models: Structural causal models represent exogenous variables, observable endogenous variables, functional assignments, and a joint distribution over exogenous variables.
- Interventions: Actions do(x) modify the treatment functions so that selected variables output fixed values, producing an intervened causal model.
- Identifiability: Identifiability requires a causal effect to be uniquely computable from the observational distribution and assumptions encoded by a causal graph.
- Selection diagrams: Selection diagrams add selection variables where structural functions or exogenous-variable distributions may differ between source and target domains.
- Selection diagrams: The absence of a selection node pointing to a variable encodes an assumption that its assigning mechanism is shared across domains.
- Transportability: Transportability means the target causal effect is uniquely computable from source observational and interventional distributions together with target observational data.
- Special cases: Direct transportability occurs when the causal relation has the same form in both domains, so no recalibration is needed.
- Non-transportability: Non-transportability requires two models agreeing on source observational, target observational, and experimental distributions while disagreeing on the target causal effect.
4 Characterizing transportable relations
Transportability is characterized through graph structures that identify when causal effects can or cannot be transferred between populations. The section moves from sC-trees to sC-forests and s-hedges, culminating in a complete graphical characterization.
- C-components are clusters of variables connected through bidirected edges and provide the structural basis for transportability analysis.A graph can be uniquely partitioned into C-components when it is not itself one.
- A Y-rooted sC-tree is a selection-confounded component whose observable nodes have at most one child, with Y descended from all nodes and selected directly.Such structures obstruct transportability; in particular, causal effects are not transportable when the graph itself is a Y-rooted sC-tree.
- For effects with X equal to Y’s parents, transportability is guaranteed when the selection diagram contains no Y-rooted sC-tree.Theorem 2 establishes this criterion for the specified class of causal effects.
- Non-transportability is not local: a W-rooted sC-tree can block transport when W is an ancestor of Y, even if the structure does not involve Y directly.Thus, transporting singleton effects is not necessarily easier than the general transportability problem.
- sC-forests generalize sC-trees by allowing a set of roots and selection nodes to point anywhere within the sC-component.Figure 4 illustrates that a causal effect may be non-transportable even without a Y-rooted or W-rooted sC-tree, because a Y-rooted sC-forest exists.
- An s-hedge is a pair of sC-forests witnessing non-transportability, with one forest intersecting X and the other not intersecting X.Theorem 4 states that when such a pair exists, the effect is not transportable; the article’s algorithm transports effects that contain no s-hedge.
5 A complete algorithm for transportability of joint effects
The sID algorithm extends identification to decide transportability by decomposing causal queries, alternating direct and trivial transport, and detecting s-hedges when transport fails.
- Algorithm: sID extends the ID algorithm to recognize transportable causal relations in selection diagrams.Its design uses c-component decomposition and hedge structures from identification theory.
- Algorithm: The algorithm partitions the input query into smaller sc-factors before attempting transport of the complete expression.This decomposition continues until the expression is transported or an s-hedge is encountered.
- Algorithm: For each sc-factor, sID alternates between direct transport and trivial transport, the latter reducing to ordinary identification.Failure produces a local graph witnessing non-transportability.
- Guarantees: The algorithm is sound: whenever sID returns an expression, that expression is correct.The soundness result is stated as Theorem 5.
- Guarantees: A causal effect is transportable if and only if the input selection diagram contains no s-hedge for the target query.The completeness corollary establishes that sID recognizes every transportable relation and identifies s-hedges as the obstruction.
6 Other perspectives on generalizability
The paper distinguishes transportability from other generalizability problems by emphasizing its combination of different populations and different study regimes, while relating it to several causal and sampling variants.
- Related problems: Classical statistical inference generalizes properties from a random sample to the population from which it was drawn.The paper presents this as a well-understood form of generalization.
- Related problems: Causal inference can be viewed as generalizing from an observational regime to an experimental regime in the same population.Imposing the experimental regime changes behavior in the population.
- Related problems: Surrogate experiments manipulate variables Z rather than X and generalize from do(Z = z) to do(X = x).The paper identifies this as a distinct causal generalization problem.
- Related problems: Selection bias generalizes from a preferentially sampled population to the population of interest.Selection depends on values of a set of selection variables.
- Distinction: Transportability concerns two distinct populations that differ in inherent characteristics and are studied under experimental versus observational regimes.This combination distinguishes it from the other generalizability problems discussed.
- Extensions: Compound-treatment transportability predicts responses under one treatment version from a study using another version.The example is exercise implemented at least fifteen minutes per day.
- Extensions: Later work extends transportability to limited source experiments and meta-transportability across multiple disparate sources.These extensions pool or restrict experimental information beyond the setting treated here.
7 Conclusions
The paper formalizes when experimental causal findings can be generalized across populations and supplies conditions and formulas for doing so, while acknowledging dependence on background knowledge about population differences.
- Motivation: Informal discussions of cross-population generalization emphasized caution but did not provide a precise formal treatment.The paper places its contribution against longstanding concerns about threats to extrapolation.
- Contribution: The article embeds generalization in mathematical language and provides theoretical guarantees when the required conditions can be ascertained.These guarantees aim to protect valid extrapolation from recognized threats.
- Contribution: A complete graphical condition decides whether target-population causal effects can be inferred, and an algorithm computes the corresponding transport formula.The framework begins with qualitative assessments of how populations differ.
- Implications: Transport formulas identify essential measurements and combine experimental and observational information into a bias-free estimand of the target quantity.The formula therefore links study design information across the two populations.
- Implications: Pooling data from both populations can increase precision in estimating commonalities and reduce sampling variability in transported relationships.The paper presents this as a potential benefit of decomposing the transport problem.
- Scope: The analysis assumes sufficient background knowledge to determine qualitatively where the two populations differ.The paper notes that such knowledge may be only partially available in practice.
Appendix 1: causal assumptions in nonparametric models
The appendix presents nonparametric structural causal models as assumption-based representations of mechanisms, with equations encoding inputs, outputs, and invariance under intervention.
- Structural causal models: The article develops its tools within nonparametric structural causal models, which unify several causal-inference approaches.SCMs are used as the framework for the transportability analysis.
- Structural causal models: An SCM represents assumptions about how the world operates rather than merely a set of probability distributions.Each functional assumption constrains the distributions the model can generate.
- Structural equations: The appendix illustrates an SCM with functions assigning X and Y from their causal inputs and exogenous variables.The example writes x = f_X(z, u_X) and y = f_Y(x, u_Y).
- Structural equations: A function’s left-hand variable is its output, while right-hand variables are inputs determining that output.An omitted variable encodes the assumption that it does not affect the output through that process.
- Interventions: The do(x) operator simulates a physical intervention by replacing the equation for X with a constant while leaving the remaining model unchanged.The resulting modified model represents do(X = x0).
Representing Interventions, counterfactuals, and causal effects
The paper represents interventions by modifying a causal model, replacing the mechanism for X with a constant while leaving other mechanisms unchanged. The resulting postintervention distribution describes outcomes under that intervention and supports comparisons across treatment levels.
- Intervention representation: The do(x) operator simulates a physical intervention by setting X to x and leaving the remaining model unchanged.The intervention is represented by deleting certain functions and replacing them with the constant X = x.
- Interventional versus observational distributions: P(z,y|do(x0)) is distinguished from the original preintervention distribution P(x,y,z).The former is the controlled or experimental distribution associated with the modified model.
- Postintervention distribution: The modified model M_x induces the postintervention distribution P_M(y|do(x)) = P_Mx(y).This distribution assigns probabilities to outcome levels under the intervened model.
- Causal-effect assessment: Treatment efficacy can be assessed by comparing aspects of the postintervention distribution at different values of x.The distribution is computed from a fully specified causal model.
Identification, d-separation and causal calculus
Identification asks whether a causal query can be estimated from observational information under stated causal assumptions. D-separation supplies graphical independence conditions, while do-calculus provides a systematic procedure for transforming interventional distributions.
- Identification: Identification asks whether P(y|do(x)) can be estimated from the observational distribution P(z,x,y) under assumptions A.In parametric models this corresponds to a unique solution for the causal quantity in terms of distributional parameters.
- Identification: A causal query is identifiable when any two models satisfying A that agree on P also agree on Q.The query then depends only on P and can be expressed using its parameters.
- Causal calculus: Do-calculus decides identifiability of do-expressions by applying inference rules whose conditions are read from the graph using d-separation.The rules map interventional and observational distributions when the relevant graphical conditions hold.
- D-separation: D-separation encodes conditional independencies that hold in every distribution compatible with the causal assumptions.It determines whether conditioning sets block paths, including paths involving colliders and their descendants.
The rules of do-calculus
The three do-calculus rules transform interventional and observational expressions through insertion or deletion of observations and actions. Repeated application aims to remove intervention operators for identification, whereas transportability requires separating them from selection variables.
- Graph transformations: The graph G_X deletes arrows entering X, while G_XZ deletes arrows entering X and emerging from Z.Z(W) contains Z-nodes that are not ancestors of W in G_X.
- Rule 1: Insertion/deletion of observations: Rule 1 permits insertion or deletion of observations when the corresponding variables are d-separated in G_X.It removes Z from P(y|do(x),z,w) under the stated graphical condition.
- Rule 2: Action/observation exchange: Rule 2 exchanges an action do(z) with observation z when the relevant d-separation condition holds in G_XZ.This maps P(y|do(x),do(z),w) to P(y|do(x),z,w).
- Rule 3: Insertion/deletion of actions: Rule 3 concerns insertion or deletion of actions in interventional distributions.The rules apply to every interventional distribution compatible with G.
- Identification versus transportability: Repeated do-calculus applications establish identifiability when the final expression contains no do-operator.For transportability, the goal instead is to separate do-operators from variables S representing population disparities.
Appendix 2
The appendix develops graphical non-transportability results and establishes completeness of the transportability procedure. It identifies sC-tree and s-hedge structures as obstacles, while showing that the relevant conditions depend on graph structure and selection-node placement.
- sC-trees: P_a(Y)(y) is transportable when G contains no subgraph forming a Y-rooted sC-tree.For direct effects, the absence of this structure supplies the stated graphical criterion.
- sC-trees: A Y-rooted sC-tree makes effects of node sets on Y non-transportable.Such structures are used to construct pairs of models agreeing on available distributions but disagreeing on the target intervention.
- Generalized obstructions: If a W-rooted sC-tree contains variables in X and W is an ancestor of Y, P_x(y) is not transportable.The result extends the obstruction from Y-rooted structures to relevant ancestors of Y.
- Graphical context: Transportability depends on the locations of X and Y and the relative positions of S-nodes, not only on the observed joint distribution.The appendix notes that the same observed distribution can require different transport formulas in different causal contexts.
- Figure 7 cases: The sp-graph and sb-graph are Figure 7 cases with no sC-tree but an sC-forest, and both are reported as non-transportable.In the sb-graph, X- and S-paths remain disjoint until intersecting at Y.
- s-hedges and completeness: A s-hedge for P_x(y) witnesses non-transportability, while the do-calculus rules are complete for establishing transportability of all such effects.The appendix states both the obstruction criterion and completeness result for the corresponding causal effects.