What is Causal Transportability?
Causal Transportability is the identification and estimation of a causal effect in a target population by combining causal evidence from a source population with assumptions and data that connect the two populations.
Quick Facts
| Specification | Official Specification |
|---|
How It Works
Separate internal validity from population relevance
Randomization can identify a source-population effect without identifying the target-population effect. A selection diagram or explicit causal graph marks mechanisms that may differ between populations and helps determine which variables must be observed in the target. Variables that predict the outcome but do not modify the treatment effect need not explain effect differences, while omitted effect modifiers can invalidate transport.
Standardize source evidence to the target distribution
Common estimators model stratum-specific outcomes or effects in the source and average them over target covariates, or use inverse odds of sampling weights to make source participants represent the target. This requires conditional exchangeability of the relevant potential-outcome contrast across populations given the chosen modifiers. Doubly robust approaches combine outcome and sampling models. Bareinboim and Pearl formalize external validity and transportability using causal diagrams.
Support and measurement define the transport boundary
Compare source and target covariate distributions, sampling scores, effective sample size, weight tails, treatment versions, outcome measurement, and temporal context. A target stratum absent from the source forces extrapolation that weighting cannot solve. Report the target estimand, transported and source estimates, uncertainty from both samples, and sensitivity to omitted effect modifiers.
Key Characteristics
- Targets a causal effect in a population distinct from the evidence source
- Separates source internal validity from target population relevance
- Requires assumptions about shared and changing causal mechanisms
- Uses target data on effect modifiers or selection variables
- Can use standardization, sampling weights, or doubly robust estimators
- Requires source-target overlap, measurement, and uncertainty diagnostics
Common Use Cases
- Transporting a randomized-trial effect to an eligible service population
- Adapting an intervention estimate across regions or deployment sites
- Reweighting an experiment to the users who will receive a policy
- Assessing whether historical causal evidence applies after population drift
- Identifying which target covariates are required before external deployment
Example
Loading code...Frequently Asked Questions
How is Causal Transportability different from internal validity?
Internal validity concerns whether the source study identifies the causal effect for its own population. Transportability concerns whether source evidence plus target data and cross-population assumptions identify the effect in a different target. A randomized source can have strong internal validity but weak transportability.
How is transportability different from generalizability?
Terminology varies, but generalizability often means extending from a study sample to its broader source population, while transportability moves evidence to a distinct target population. The operational requirement is to define source, target, estimand, and selection process explicitly rather than rely on the label.
Which variables should be adjusted for when transporting an effect?
Adjustment should cover measured variables sufficient to explain relevant source-target differences, especially treatment-effect modifiers and selection variables indicated by the causal structure. Pure outcome predictors do not necessarily need balancing if they do not modify the effect or participate in required identification paths.
Can weighting solve a lack of target-population overlap?
No. Sampling or inverse-odds weights can rebalance strata represented in both populations. If a target subgroup or treatment version has no source support, transport requires extrapolation, additional evidence, a narrower target population, or an explicit structural assumption that must be defended.
How should a transported causal effect be validated?
Audit source identification first, then compare population definitions, covariate measurement, sampling-score overlap, weight tails and effective sample size. Report source and transported estimates with uncertainty, examine known target outcomes when available, and test sensitivity to omitted effect modifiers and mechanism changes.