Why Support Gaps Require a Decision
Transporting a causal effect means using information from a source population to learn about an effect in a target population. A central requirement is that the source contain relevant information for the covariate profiles that occur in the target. When some target profiles have no source counterparts, the issue is not merely that an estimator may be imprecise: the source data do not directly reveal the conditional causal effect in those profiles.
There is no single automatic repair. An analyst can restrict the target population to profiles represented in the source, define a different estimand whose covariate distribution lies within source support, or extrapolate with an outcome model. These choices answer different questions and rely on different assumptions. The choice should be made explicit rather than hidden inside a weighting or prediction procedure.
For a binary treatment, the target average treatment effect (ATE) is \(\psi_T=E\{Y(1)-Y(0)\mid S=0\}\). Under the usual identification conditions for transport, it can be written as an average of source-identifiable conditional effects:
where \(f_T(x)\) is the target covariate distribution. For discrete \(X\), the integral is a sum. This expression requires, among other conditions, that conditional effects are transportable given \(X\), and that the source provides information to identify \(\tau(x)\) wherever the target places positive probability. The latter is the relevant support condition: if \(f_T(x)>0\) but the source has no observations at \(x\), the source data alone do not identify \(\tau(x)\) there.
A support gap differs from sparse overlap. If the source has some observations near a target profile, the effect may be identifiable in principle but estimated imprecisely; weighting can also become unstable. If the source has no observations in a region, a purely data-based estimate of the conditional effect there is unavailable. More data from the same source distribution may reduce sampling error in supported regions, but cannot fill a structural gap.
What Can Be Identified Without Extrapolation?
Let \(q=P(X\in\mathcal{O}\mid S=0)\), the proportion of the target population in common support. If \(q>0\), the target ATE can be decomposed into its supported and unsupported parts:
This follows by partitioning the target expectation according to whether \(X\) is in \(\mathcal{O}\) and applying the law of total expectation. The first term can be estimated by standardizing conditional effects over the target covariate distribution within common support, if the identification assumptions hold there. The second term cannot be learned from source data without additional information or assumptions. If \(q=1\), this particular support obstacle is absent; if \(q<1\), reporting the full target ATE requires a strategy for the unsupported contribution.
Worked Example: Restricting the Target Population
Suppose a hypothetical target health program includes 1,000 adults. The source trial includes adults aged 50–69 but none aged 70 or older. In the target, 400 adults are aged 50–59, 200 are aged 60–69, and 400 are aged 70 or older. The trial estimates treatment effects of 3 and 5 additional symptom-free days in the first two age groups. Assume treatment is randomized in the trial and, within each age group, its conditional effect transports to the target.
The full target includes 40% of adults aged 70 or older, a group with no source support. Instead, define the restricted target as target adults aged 50–69. Its age distribution is \(400/600\) aged 50–59 and \(200/600\) aged 60–69. Standardizing the two conditional effects to that restricted distribution gives:
The estimate describes an average gain of about 3.67 symptom-free days for target adults aged 50–69, under the stated transport assumption. It is not an estimate for all 1,000 target adults. The restriction makes the supported population the estimand; it does not establish what the effect is among adults aged 70 or older.
Three Responses to a Support Gap
The available responses differ in whether they change the population being described or make stronger assumptions about unsupported outcomes. The restriction in the example above answers a question about a literal target subgroup. A redefined estimand instead specifies another covariate distribution for standardization. An outcome-model approach retains the original target but predicts effects in unsupported regions.
| Strategy | What is estimated? | Main cost or assumption |
|---|---|---|
| Restrict the target | Effect among target units in a supported subgroup | Does not answer the full-target question |
| Redefine the estimand | Effect standardized to a specified distribution with source support | The new population or weighting distribution must be substantively meaningful |
| Extrapolate with an outcome model | Effect for the original target, including unsupported profiles | Requires the model’s conditional outcome predictions to be valid beyond observed source support |
Redefining the Estimand
A standardized effect is determined not only by conditional effects but also by the covariate distribution over which they are averaged. Instead of the original target distribution \(f_T(x)\), an analyst can define a reference distribution \(q(x)\) that assigns probability only to covariate values with source support:
This is a well-defined causal estimand if its population and treatment contrast are clearly specified. It may describe a policy-relevant reference population or a chosen mix of supported profiles. It is not automatically the ATE in the original target. In particular, choosing \(q\) because it is convenient does not make \(\psi_q\) a target-population effect.
Worked Example: Standardizing to a Supported Reference Population
Consider a hypothetical target with three covariate groups: \(L\), \(M\), and \(H\), with target proportions 0.50, 0.30, and 0.20. The source contains groups \(L\) and \(M\), but no members of \(H\). Suppose the conditional treatment effects identified and transported from the source are 2 units in \(L\) and 6 units in \(M\). To avoid asserting an effect for \(H\), define a reference estimand using weights \(q(L)=0.25\), \(q(M)=0.75\), and \(q(H)=0\). The weights are nonnegative and sum to one.
The resulting standardized effect is:
The value 5.00 is the average effect for the specified reference mix of supported groups. It is not the target ATE: the target assigns 20% probability to group \(H\), while \(q\) assigns none, and the \(L\)-to-\(M\) mix also differs from the target's. The choice of \(q\) should therefore be justified by the question of interest, not presented as a technical workaround that preserves the original estimand.
Extrapolating with an Outcome Model
An outcome model can generate predictions for unsupported covariate values. For example, it can model the conditional mean outcome under each treatment level, \(m_a(x)=E(Y\mid A=a,X=x,S=1)\), and use the predicted difference \(m_1(x)-m_0(x)\) in the target average. This approach can retain the original target estimand, but predictions outside source support are extrapolations: they follow from the model form rather than direct source comparisons at those covariate values.
For a causal interpretation, the model does not replace the identification conditions. Treatment assignment must be adequately controlled or randomized in the source, the relevant conditional effects must transport given the covariates, and the outcome-model predictions must be correct enough for the target profiles. Most importantly, an extrapolation claim requires a defensible model for how outcomes or effects behave beyond the observed source range. A good fit within that range does not, by itself, verify that extension.
Worked Example: Extrapolating a Linear Treatment Effect
Suppose a hypothetical randomized source study measures a continuous covariate \(X\), observed from 0 to 1, while the target has \(P(X=0)=0.40\), \(P(X=1)=0.30\), and \(P(X=2)=0.30\). Consider fitted source outcome models, in outcome units, of \(m_0(x)=10+x\) and \(m_1(x)=12+2x\). Assume randomization identifies the source conditional means over its observed range, and suppose the analyst is willing to assume these linear mean models remain valid through \(x=2\).
The predicted conditional treatment effect is:
Thus, the model predicts effects of 2 units at \(x=0\), 3 units at \(x=1\), and 4 units at \(x=2\). Standardizing these predictions to the target distribution gives:
The contribution \(0.30(4)=1.20\) comes from the target group at \(x=2\), which is outside source support. It is a model-based prediction, not a conditional effect directly estimated from source observations at \(x=2\). If the true effect at \(x=2\) were, for example, 1 rather than 4, the target average would be lower by \(0.30(3)=0.90\) units. The numerical result is therefore conditional on the extrapolation assumption, and uncertainty from that assumption should not be obscured by reporting only a standard error from the fitted model.
A Practical Decision Process
A defensible response starts by locating the gap and then stating which question the analysis will answer. The following sequence separates the support diagnosis from the estimand decision.
Identify which covariate profiles occur in the target but not the source, and estimate how much target probability lies there. Distinguish no source support from a small number of source observations.
Decide whether the scientific question concerns all target members, a supported target subgroup, or a clearly specified reference distribution.
Restrict standardization to a supported target subgroup, define and report a new standardization distribution, or use an outcome model whose extrapolation assumptions can be defended.
For a restricted or redefined estimand, say which target population it represents. For extrapolation, identify which target regions rely on predictions rather than source data.
Common Mistake
Wrong reasoning: “The outcome model predicts an effect for every target covariate value, so the full target ATE is identified.” A prediction can be computed wherever a model returns a number, but computation alone does not establish identification or validate the model outside source support.
Correction: Separate the estimand from the assumptions used to estimate it. Restricting or redefining the estimand avoids needing effects in unsupported regions but changes the population being described. Keeping the full target estimand through extrapolation requires explicit assumptions about outcomes or effects in those regions, and the resulting conclusion should be reported as model-dependent.
Check Your Understanding
Use the distinction between the target population, the estimand, and the assumptions to answer these questions.
- What is the difference between a support gap and sparse overlap?
- If 25% of the target population lies outside source support, which part of the target ATE cannot be learned from source data without further assumptions?
- How does the interpretation change when the target population is restricted to a supported subgroup?
- Why is an effect standardized to a reference distribution \(q(x)\) not automatically the original target ATE?
- In the linear-model example, which target contribution depends on extrapolation, and what assumption makes that calculation possible?