Transporting Treatment Effects from Randomized Trials to Target Populations

Randomized trials are designed to estimate treatment effects with strong internal validity. Their participants, however, are often selected through eligibility criteria, referral pathways, geography, recruitment practices, and willingness to participate. The resulting trial effect may not equal the effect that matters for a broader health system, a national population, an underserved community, or a future implementation setting.

Transportability analysis addresses this external-validity problem by making the target population explicit and using information about differences between the trial and target data. The goal is not to declare a trial “representative” in general. It is to estimate a defined target-population treatment effect under stated causal assumptions and to show how much the answer depends on the available covariates, overlap, and model specification.

1. Start with the target estimand

A treatment effect has no single universal value outside its population and outcome definition. Before analyzing transportability, specify the intervention, comparator, eligible population, treatment assignment strategy, outcome, follow-up period, contrast measure, and target population. A target average treatment effect may be a risk difference at 12 months, a risk ratio, an odds ratio, a survival contrast, or another prespecified summary.

Use an estimand framework to describe what happens when participants deviate from assigned treatment, receive rescue therapy, die before a later outcome, or become lost to follow-up. These intercurrent events can affect both the trial estimand and the target-population estimand. Transporting a poorly defined treatment effect does not solve an estimand problem; it moves the ambiguity to a new population.

The target population may be a sample from the same source population as the trial, a broader population containing trial participants, or a distinct population that does not overlap with the trial. This distinction determines whether the problem is usually described as generalizability or transportability. The data structure and identification strategy must match that distinction.

2. Generalizability and transportability are related but distinct

Generalizability commonly refers to extending a study result from a study sample to the population from which that sample was drawn. The target may contain the trial participants as a subset. Transportability usually refers to moving an effect from a study sample to a different target population, such as people in another health system, region, or eligibility frame.

The terminology is less important than stating the data relationship clearly. Readers need to know whether trial and target data overlap, which covariates are available in both sources, and whether the target data represent the population to which a decision will be applied. A target database assembled from healthcare users may not represent all eligible residents, and a census-like data source may not contain the clinical variables needed to identify effect modifiers.

External validity should be treated as a design and analysis objective. ICH E9 emphasizes that the basis for generalization to the intended patient population should be understood and explained, including how trial centers and populations support the inference. Modern transportability methods provide a quantitative framework for making that reasoning explicit.

3. Identify effect modifiers, not every baseline variable

Transportability depends on differences in variables that modify the treatment effect. A variable can strongly predict the outcome yet be irrelevant to treatment-effect heterogeneity. Conversely, an interaction between a pretreatment characteristic and treatment may matter for transport even when the characteristic is not a major prognostic factor.

Construct a causal diagram or a prespecified scientific rationale for selecting candidate effect modifiers. Examples may include baseline disease severity, prior treatment, age, comorbidity, disease stage, biomarker status, or care context when they plausibly alter the relative or absolute response to treatment. Avoid adjusting for post-treatment variables, mediators, or colliders merely because they differ between trial and target data.

Selection should balance causal relevance and measurement quality. A theoretically important effect modifier that is defined differently across sources can introduce more uncertainty than it resolves. Document variable definitions, units, timing, missingness, and harmonization rules before estimating sampling weights or standardization models.

4. The core identification assumptions

Transporting a treatment effect requires assumptions beyond randomization within the trial. First, the trial must identify the causal effect for its enrolled participants under the chosen treatment strategy. Second, conditional exchangeability or conditional transportability requires that, after controlling for the selected pretreatment covariates, treatment-effect differences between the trial and target populations are explained by those covariates.

Third, positivity requires adequate representation of target covariate patterns in the trial, or trial covariate patterns in the target data, depending on the design. If a target subgroup is absent from the trial, no weighting method can recover a treatment effect there without additional assumptions or external evidence. Extreme weights are often a diagnostic signal of limited overlap rather than a nuisance to be trimmed silently.

Fourth, the outcome and covariate measurements must be sufficiently comparable across data sources. A target population with a different outcome ascertainment process may appear to have a different treatment effect even when the biological response is similar. Finally, the models used for sampling, outcome response, or effect modification must be adequately specified, or the analyst must use robust or flexible approaches with honest uncertainty assessment.

5. Standardization: predict trial effects in the target

Standardization, also called the g-formula approach in this setting, estimates treatment-specific outcomes across covariate patterns and averages those predictions over the target population distribution. In a simple implementation, fit an outcome model using the randomized trial, predict the outcome under each treatment for every target individual, and average the predictions.

This approach is intuitive because it directly answers the question: what would the target population’s average outcome be under treatment A or treatment B if the trial-based conditional outcome relationships applied? It also makes the role of effect modifiers visible. If the target distribution places more weight on a subgroup with a different treatment effect, the transported average can differ from the trial average.

Standardization is vulnerable to outcome-model misspecification. Flexible regression, targeted learning, or doubly robust estimators may reduce reliance on one model, but they do not eliminate the need for valid measurements, overlap, and a credible set of effect modifiers. Report the model form, interactions, transformations, regularization, and uncertainty procedure.

6. Sampling weights: reweight the trial to the target

Sampling-weight approaches estimate how likely each individual is to belong to the trial rather than the target population, conditional on covariates. Trial participants who resemble common target profiles receive more weight; trial participants with profiles overrepresented in the trial receive less relative weight. The weighted trial can then estimate a target-population treatment effect.

For generalization, inverse probability of sampling weights may be used when the study sample is a subset of the target population. For transportability to a distinct target, Westreich and colleagues describe inverse odds of sampling weights. The distinction matters because the target and trial populations may be nonoverlapping, and the weighting formula must reflect the study design.

Inspect the weight distribution, effective sample size, covariate balance, and positivity. A small number of extreme weights can make the transported effect unstable and inflate variance. Stabilization or truncation may be considered, but the thresholds and their effect on the estimand must be prespecified and reported rather than chosen only to produce a preferred result.

7. Doubly robust and augmented estimators

Doubly robust estimators combine a model for trial participation or sampling with a model for the outcome or treatment effect. Under suitable conditions, consistency can be retained if one of the two models is correctly specified, although this property does not protect against poor overlap, measurement error, or violations of conditional transportability.

Augmented inverse probability weighting can improve precision when outcome information is informative, but the implementation requires careful attention to the trial design, target sampling fraction, missing data, and treatment-specific predictions. Researchers should report which data contributed to each component and how standard errors were calculated.

Machine-learning models can be used for nuisance functions, but algorithmic flexibility does not turn an unsupported transport assumption into evidence. Cross-fitting, regularization, calibration, and sensitivity analysis remain important, particularly when the number of effect modifiers is large relative to the target or trial sample.

8. Evidence summary table

9. A practical trial-to-target workflow

10. Diagnose overlap and effective sample size

Overlap is a practical condition for learning about the target population from the trial. Plot the distribution of participation probabilities or sampling weights by treatment group and target-relevant subgroup. Compare covariate distributions before and after weighting, and report the effective sample size of the weighted trial.

When overlap is weak, the transported estimate may be driven by extrapolation. Trimming observations can improve stability but changes the target population or estimand. It should be described as a scientific restriction, not merely a technical cleanup. Alternative strategies include redefining the target population to the region of common support, collecting trial data in underrepresented profiles, or using external evidence to inform effect modification.

Positivity should be assessed for the variables that modify treatment effect, not only for variables that predict participation. A model can achieve excellent balance on sampling predictors while failing to represent a clinically important treatment-effect modifier.

11. Sensitivity analysis for unmeasured effect modification

No observed-data method can verify that all treatment-effect modifiers have been measured. Sensitivity analysis should therefore address how a plausible unmeasured modifier could change the transported effect. Specify its prevalence or distribution, association with treatment response, and difference between trial and target populations.

Vary the strength of effect modification and the degree of imbalance to identify tipping points. Report whether the clinical conclusion changes across plausible scenarios. This is more informative than stating that the analysis is limited by unmeasured confounding when the central issue is unmeasured effect modification between trial and target populations.

Sensitivity analyses can also vary outcome definitions, target-population inclusion criteria, missing-data assumptions, sampling models, weight truncation rules, and treatment-effect scales. The analysis should preserve the target question while identifying which assumptions are carrying the inference.

12. Common interpretation errors

A transported effect is not automatically the “real-world effect.” It is an estimate for a specified target under stated assumptions. Calling a target database representative without describing its sampling frame, care access, outcome ascertainment, and covariate harmonization is not a substitute for an external-validity analysis.

Another error is transporting a marginal effect without considering effect modification. If treatment effects vary across baseline risk or disease severity, the target distribution can change the average effect even when treatment is randomized in the trial. A third error is ignoring clinical implementation. Treatment availability, adherence, monitoring, competing therapies, and patient preferences may differ in the target setting and can change the practical effect of an intervention.

Finally, do not report a narrow confidence interval as proof that the target-population inference is secure. Sampling uncertainty is only one component of uncertainty. Model dependence, measurement harmonization, overlap, and unmeasured effect modification may dominate the decision.

Conclusion

Transporting treatment effects from randomized trials to target populations turns external validity into an explicit causal-inference problem. Define the target estimand, identify effect modifiers, harmonize trial and target data, evaluate positivity and conditional transportability, and use standardization or sampling weights with transparent sensitivity analyses. The result is not a universal correction for an unrepresentative trial; it is a target-specific estimate whose credibility depends on design, measurement, overlap, and assumptions. Reporting those conditions clearly allows clinicians, regulators, and policy researchers to judge whether the evidence supports decisions in the population that matters.