03 · What You Need to Know
How Causal Findings Can Be Extended Across Populations
What Is Transportability in Research?
Transportability concerns whether causal knowledge established in one population or environment can be used to infer a causal effect in another population or environment.
For example, researchers may evaluate a learning intervention among university students with extensive digital access and seek to estimate its effect among students in institutions with more limited technological resources.
The task is not simply to repeat the original estimate in a new report. It requires determining whether differences between the source and target populations affect the causal quantity being estimated.
In formal causal inference, transportability may involve combining evidence from a source experiment with information about the target population under explicit assumptions.
Some transportability problems can be addressed using measured effect modifiers and population distributions. Others require stronger structural assumptions or additional data.
Transportability is therefore an inferential problem rather than a statistical technique that can be applied automatically to any study.
What Is Generalizability?
Generalizability broadly concerns extending research findings beyond the observations directly studied.
In causal inference literature, a narrower distinction is sometimes made between generalizing from trial participants to a broader population that includes them and transporting findings to a separate population.
For example, a trial may enroll a subset of eligible students from a university system. Generalization may concern the average treatment effect among all eligible students in that system.
Transportability may instead concern students in another educational system who were not part of the original source population.
However, authors sometimes use generalizability as an umbrella term that includes both situations.
Researchers should therefore define the population relationship rather than assume that terminology alone establishes the methodological distinction.
How Do the Two Concepts Compare?
| Dimension |
Generalizability |
Transportability |
| Typical population relationship |
Study participants to a broader population containing or representing them |
Source population to another, potentially separate target population |
| Central question |
Does the finding apply to the intended broader population? |
What effect would occur in the different target population? |
| Relevant evidence |
Sampling, participation, effect modifiers, population data |
Source causal evidence, target-population information, structural assumptions |
| Common analytical approaches |
Standardization, weighting, outcome modeling |
Standardization, weighting, transport formulas, causal modeling |
| Shared limitation |
Unmeasured effect modification or insufficient overlap |
Unmeasured effect modification, insufficient overlap, or changed causal mechanisms |
The distinction is useful for clarifying the task, but the underlying assumptions matter more than the label.
Why Is Internal Validity Not Enough?
Internal validity concerns whether the study provides a credible estimate of the causal effect within its design and population.
Randomization can support treatment-group comparability in expectation, but it does not make trial participants representative of every population in which the intervention might be used.
Suppose the original trial estimates an average treatment effect of eight examination points.
That estimate may be internally valid for the studied population while differing from the average effect in another population with different distributions of relevant effect modifiers.
Researchers should distinguish credible within-study causal inference from broader applicability.
Transportability generally builds on credible source evidence. Transporting a biased effect estimate does not remove its original bias.
What Is a Target-Population Average Treatment Effect?
A target-population average treatment effect, or target ATE, is the average causal contrast between two specified treatment conditions among units in the target population.
Using potential-outcome notation, it may be expressed as:
To estimate this quantity using a source study, researchers need assumptions connecting the source evidence to the target population.
The target population must also be specified precisely. A target consisting of students at one university is different from a target consisting of all undergraduate students in a national higher education system.
Why Are Effect Modifiers Central to Transportability?
An effect modifier is a characteristic across which the causal effect of an intervention differs on a specified effect scale.
For example, connectivity may modify the effect of a digital learning platform if students with reliable internet access receive more of the intended instructional activities.
If the source and target populations have different connectivity distributions, their average treatment effects may differ.
However, not every characteristic associated with outcomes is necessarily an effect modifier.
Prior achievement may predict examination performance without changing the intervention's causal effect. Conversely, a characteristic with a modest relationship to baseline outcomes may substantially modify treatment effects.
Researchers should identify plausible effect modifiers through substantive theory, previous evidence, and knowledge of the intervention.
Transportability requires attention to variables that make the source and target causal effects differ, not simply a mechanical comparison of every demographic characteristic.
What Is Conditional Exchangeability Across Populations?
A common identifying assumption is that, after conditioning on an appropriate set of measured characteristics, the relevant potential outcomes or treatment effects can be related across source and target populations.
For transporting an average treatment effect, a sufficient condition under some designs is that the conditional average treatment effect is the same in source and target populations for each relevant covariate profile.
This is sometimes described as conditional exchangeability in measure between populations.
It is weaker than requiring the entire potential-outcome distributions to be identical, but it remains a substantive assumption.
For example, researchers might assume that among students with comparable digital proficiency and connectivity, the platform's average effect is the same across two educational systems.
That assumption may be questionable if curricula, instructional support, language, or intervention delivery differ in consequential ways.
Observed data can inform its plausibility, but the required cross-population counterfactual relationships generally cannot be verified conclusively from the source trial alone.
Why Is Positivity or Population Overlap Necessary?
Positivity concerns whether the source data provide support for the relevant characteristics represented in the target population.
Suppose the target population includes students with very limited connectivity, but every participant in the source trial has reliable internet access.
The source trial provides no direct evidence about the intervention's effect in the excluded connectivity category.
Weighting cannot create treatment-effect information for an entirely unobserved subgroup.
Transporting the effect to that subgroup would require extrapolation, additional evidence, or stronger modeling assumptions.
Even when overlap is technically present, very small source samples for important target subgroups may produce unstable estimates.
Researchers should therefore examine both structural overlap and the practical amount of information available.
How Does Standardization Transport a Treatment Effect?
Under suitable identification assumptions, researchers may estimate treatment effects within levels of measured effect modifiers in the source study and average those effects using the target population's distribution.
For a discrete covariate X, a simplified expression is:
The expression illustrates how a target effect may differ from the source average when populations contain different proportions of relevant subgroups.
It is not a universal transport formula for every causal structure. More complex problems may involve additional variables, selection mechanisms, or interventions that differ across environments.
Can Weighting Be Used Instead of Standardization?
Yes. Weighting methods may reweight source-study participants so that their measured characteristics better reflect the target population.
Depending on the design, researchers may estimate study participation probabilities or density ratios between source and target covariate distributions.
These weights can support target-population effect estimation under appropriate assumptions.
However, extreme weights may increase variance, and unmeasured effect modifiers can remain problematic.
Weighting and standardization are alternative or complementary estimation strategies. Neither establishes the identifying assumptions merely by producing a numerical estimate.
Researchers should also distinguish weights used for transporting treatment effects from ordinary survey design weights.
Can an Effect Be Transported When the Intervention Changes?
Not automatically. Transportability usually requires a sufficiently clear definition of the treatment strategies being compared.
Suppose the source trial evaluates a platform accompanied by weekly instructor guidance, while the target institution intends to deploy the platform without that guidance.
The target intervention may differ from the source intervention in a causally important way.
Even if the populations were otherwise similar, the original effect estimate might not represent the effect of the modified strategy.
Researchers should distinguish population differences from changes in the intervention itself.
When an adaptation changes active components, new evidence or assumptions about the adapted intervention may be necessary.
How Does Transportability Differ From Replication?
Replication collects new evidence to examine whether findings persist under specified conditions.
Transportability uses existing causal evidence, often together with target-population information, to infer an effect in another population under stated assumptions.
A new randomized trial in the target population may provide more direct evidence than transporting the original estimate, although it may be expensive or impractical.
Replication and transportability can complement each other. Target-population studies may test assumptions, reveal previously unrecognized effect modifiers, or refine transport models.
However, agreement between two estimates does not prove that all transport assumptions are correct, particularly when both studies share limitations.
Can Transportability Apply to Qualitative Research?
The term transportability is most formally developed in causal inference, particularly for extending causal effects across populations or environments.
Qualitative researchers more commonly discuss transferability, analytical generalization, or theoretical relevance.
These approaches can examine whether mechanisms, interpretations, or experiences may be relevant elsewhere, but they do not ordinarily estimate a target-population average treatment effect.
Researchers should avoid using the terms as interchangeable when their inferential purposes differ.
The broader issue of extending findings without statistical representativeness includes several methodological traditions that require distinct justifications.
What Can Make a Transportability Claim Unconvincing?
A transportability claim is weak when the target population is poorly defined, important effect modifiers are unmeasured, relevant target subgroups are absent from the source data, or intervention conditions differ without adequate justification.
It may also be undermined by bias in the source study or unreliable target-population data.
For example, a randomized trial may estimate a valid effect among participating students but collect no information about digital access, even though digital access plausibly modifies the effect.
If the target population differs substantially in digital access, the researchers may lack information needed for credible transport.
Researchers should not interpret a successful statistical model fit as proof of external validity.
How Should Transportability Be Reported?
Reports should identify the source and target populations, intervention and comparison conditions, target estimand, relevant covariates, and assumptions supporting transport.
Researchers should describe the target-population data, the estimation method, overlap diagnostics, and sensitivity to plausible alternative assumptions.
Where important contextual conditions differ, explain why the causal effect is expected to remain comparable or how those differences were addressed.
It is also useful to distinguish an effect directly estimated in the source study from an effect estimated for the target population through modeling.
This transparency allows readers to evaluate whether the proposed extension is supported by evidence rather than terminology alone.
