Why the Choice of Covariates Matters
A study and its target population can differ in age, disease severity, clinical setting, or other characteristics. Those differences do not automatically prevent transport: what matters is whether they change the causal effect being transported, and whether the effect-relevant characteristics are distributed differently across populations. A variable can strongly predict the outcome without changing the treatment effect. Treating every outcome predictor as a required transport variable can therefore obscure the key question.
The first step is to specify the target causal contrast and its scale. This tutorial uses the average treatment effect on the risk-difference scale for a binary outcome. Let \(A\in\{0,1\}\) denote treatment, \(Y^a\) the outcome that would be observed under treatment level \(a\), and \(S=1\) denote membership in the study source population, with \(S=0\) denoting the target population. For a covariate vector \(X\), the conditional treatment effect in population \(s\) is the difference in conditional mean potential outcomes. A variable modifies the effect when this difference varies across its values.
The distinction is causal, not merely a matter of whether a regression coefficient is statistically significant. A strong predictor can have the same association with the outcome in both treatment groups, leaving the treatment contrast unchanged. Conversely, a modest predictor can modify the treatment effect if the contrast differs across its levels.
Standardizing Conditional Effects to the Target
Suppose \(Z\) is a set of measured covariates chosen so that, within levels of \(Z\), the causal effect is stable between source and target. Write the conditional source effect as \(\delta_1(z)=E[Y^1-Y^0\mid Z=z,S=1]\). The target average treatment effect can then be obtained by averaging those conditional effects over the target distribution of \(Z\), provided the conditional effects transport and all needed \(Z\) values are represented in the source.
For continuous \(Z\), the sum is replaced by an integral. The equality follows by conditioning the target effect on \(Z\): if \(E[Y^1-Y^0\mid Z=z,S=0]=\delta_1(z)\), then averaging that conditional effect over the target distribution gives the target marginal effect. This is a transport condition for the causal contrast; it does not require the source and target to have the same distribution of \(Z\).
This argument also clarifies why choosing covariates is not simply a search for variables associated with the outcome. If a variable predicts both potential outcomes but leaves their difference unchanged, changing its distribution between source and target does not change the transported risk difference. If it modifies the risk difference and its distribution differs, ignoring it can yield the wrong target effect.
Worked Example: A Strong Prognostic Factor That Does Not Modify the Effect
Consider a hypothetical randomized study of a treatment for a binary outcome. Let \(X=1\) indicate severe baseline illness. Suppose the source has 20% severe patients and the target has 80%. The outcome risks are as follows.
| Illness severity | Risk under control, \(Y^0\) | Risk under treatment, \(Y^1\) | Risk difference, \(Y^1-Y^0\) |
|---|---|---|---|
| Not severe, \(X=0\) | 0.10 | 0.00 | -0.10 |
| Severe, \(X=1\) | 0.30 | 0.20 | -0.10 |
Severity is prognostic: under control, risk is 0.30 for severe patients and 0.10 for others. But the treatment reduces risk by 0.10 in both groups, so severity does not modify the risk difference. In the source, the marginal control risk is \(0.80(0.10)+0.20(0.30)=0.14\), and the treatment risk is \(0.80(0.00)+0.20(0.20)=0.04\). The source risk difference is \(0.04-0.14=-0.10\).
In the target, the corresponding control risk is \(0.20(0.10)+0.80(0.30)=0.26\), and the treatment risk is \(0.20(0.00)+0.80(0.20)=0.16\). The target risk difference is \(0.16-0.26=-0.10\). The target risks differ from the source risks because severity is distributed differently, but the risk difference is unchanged. Thus severity is not needed to transport this risk difference under the stated data-generating pattern, even though it is important for predicting outcome risk.
When Effect Modification Changes the Transported Effect
Now suppose a covariate changes the magnitude of the treatment effect, and its distribution differs across populations. Then the source marginal effect averages subgroup effects using source proportions, while the target effect must use target proportions. Applying the source marginal effect directly to the target would use the wrong mixture.
Worked Example: Standardizing Over a Genuine Effect Modifier
In a hypothetical randomized study, let \(Z=1\) indicate a biological marker associated with a larger treatment benefit. The source population is 25% marker-positive, while the target population is 75% marker-positive. Risks in the two marker groups are:
| Marker status | Risk under control | Risk under treatment | Risk difference |
|---|---|---|---|
| \(Z=0\) | 0.20 | 0.10 | -0.10 |
| \(Z=1\) | 0.40 | 0.10 | -0.30 |
The risk difference varies by marker status, so \(Z\) is an effect modifier on this scale. The source marginal risk difference is
To transport the effect, use the target proportion of marker-positive individuals:
Under the assumed stable subgroup effects, the transported target risk difference is \(-0.250\), or 25 percentage points fewer outcome events under treatment. The source marginal effect is only \(-0.150\), or 15 percentage points fewer events, because the source contains fewer marker-positive individuals. The calculation assumes that both marker groups occur in the source and that the conditional effects apply to the corresponding target groups.
The formal transport condition depends on the estimand. For a target risk difference, equality of conditional risk differences across source and target is sufficient for the standardization above. Transporting each potential-outcome risk separately generally requires a stronger condition: the conditional mean of each potential outcome must be comparable across populations given the selected covariates. A variable can therefore be unnecessary for transporting one contrast but relevant for transporting arm-specific risks or another estimand.
A Principled Strategy for Selecting Variables
Selection should start from the causal question and the population difference, not from a list of the strongest outcome predictors. A defensible covariate set should make the conditional effect transport assumption plausible while remaining measurable and supported in the target and source. No finite dataset can prove that a candidate variable is not an effect modifier, so substantive knowledge and transparent sensitivity analysis remain important.
Specify the treatment contrast, outcome, target population, and effect scale. A variable may modify a risk difference but not modify another contrast, so “effect modifier” is incomplete without a scale.
Identify variables needed to control treatment-outcome confounding within the source, where relevant. That role is distinct from being needed to explain effect differences between source and target. A variable can serve both roles, either role, or neither.
Use subject-matter knowledge and the causal structure to identify characteristics associated with changes in the causal response across populations. Then ask whether those characteristics, or their causes, have different distributions in source and target.
Evaluate whether candidate covariates plausibly alter the treatment contrast, not merely whether they predict the outcome. Main-effect associations are not evidence by themselves that a variable must enter the transport standardization.
The needed covariates must be measured comparably and available for the target distribution used in standardization. The source must contain information for the covariate values to which its conditional effects are applied; otherwise the proposed transport calculation is not supported by the data.
Selection diagrams and the earlier material on conditional exchangeability across study participation provide a formal way to organize these judgments. In practical terms, the question is whether conditioning on the chosen covariates makes the relevant conditional causal effect independent of study participation. A covariate that causes selection into the study is not automatically an effect modifier; it matters for transporting the contrast when it carries, or helps explain, a difference in conditional causal effects.
Worked Example: Choosing Between a Prognostic Covariate and a Candidate Modifier
Suppose a hypothetical trial evaluates a preventive treatment. Two measured variables are available in both the trial and the target population. \(W\), age group, predicts outcome risk, but clinical evidence suggests the treatment risk difference is the same across age groups. \(V\), a marker of the biological pathway targeted by treatment, is expected to change the treatment response, and its prevalence is 30% in the trial but 70% in the target. The trial is randomized, and both marker groups are represented.
The target estimand is the average risk difference. Because randomization identifies the trial’s treatment contrast without adjustment for baseline confounding, neither \(W\) nor \(V\) is required for that particular treatment-confounding role. That does not settle their transport roles.
For \(W\), outcome prediction is not enough to require standardization. If the conditional risk difference is constant over age and the target and trial share that conditional effect, a change in age composition alone does not change the target risk difference. \(W\) may still be useful for precision or for transporting absolute risks, but those are separate reasons.
For \(V\), the scientific expectation of effect modification and the substantial prevalence difference make it a central transport variable. The analyst should estimate trial effects by marker status, examine their uncertainty, and standardize those conditional effects to the target prevalence. If the estimated risk differences were \(-0.08\) for \(V=0\) and \(-0.24\) for \(V=1\), the target estimate would be \(0.30(-0.08)+0.70(-0.24)=-0.024-0.168=-0.192\). The arithmetic alone does not establish that these conditional effects are transportable; that requires the substantive assumption that effects within marker status apply in the target.
Evidence, Uncertainty, and Common Pitfalls
Treatment-by-covariate interaction estimates can help assess effect modification, but a nonsignificant interaction test does not establish that effects are identical. Such tests may have low power, especially when subgroup sizes are small. Conversely, trying many interactions can produce unstable patterns. A careful analysis therefore combines prior scientific knowledge, causal structure, observed conditional effects, their uncertainty, and how different the source and target distributions are.
A useful distinction is between a variable being an effect modifier and being necessary in a particular transport calculation. If a modifier has the same distribution in source and target, the source and target marginal effects can agree despite within-population heterogeneity. If a modifier differs in distribution, failing to account for it can change the target average. The practical priority is especially high when both effect heterogeneity and population-composition differences are plausible.
- Pitfall: selecting every strong outcome predictor. The mistaken argument is, “This variable predicts the outcome, so it must be adjusted for to transport the effect.” Prediction does not imply modification. For a risk difference, a prognostic variable with the same conditional treatment effect across its levels need not be included to correct a composition-induced change in the marginal effect.
- Pitfall: using only statistical significance to select modifiers. The mistaken argument is, “The interaction test was not significant, so this covariate cannot modify the effect.” A failure to reject may reflect imprecision. Report uncertainty and use scientific knowledge rather than treating a \(p\)-value threshold as proof of invariance.
- Pitfall: confusing transport adjustment with confounding control. A covariate required to control treatment assignment in an observational source study is not automatically a transport modifier. Conversely, randomization does not remove the need to address effect modification across populations.
Check Your Understanding
Answer each question using the risk-difference scale unless another scale is specified.
- A covariate predicts the outcome under both treatment levels, but the risk difference is constant across its values. Is it an effect modifier on this scale? Explain.
- Why can a source marginal treatment effect differ from the target effect when a covariate modifies effects and has different source and target distributions?
- State the standardization formula for a binary covariate \(Z\) when its conditional effects are transportable from source to target.
- Why does a nonsignificant treatment-by-covariate interaction test not establish that the covariate is not an effect modifier?
- Give one reason a covariate may be needed for estimating an effect within an observational source study and explain why that reason is distinct from its transport role.