Tutorials › Causal Inference › Transport from Aggregate Target Covariate Margins

Transportability and Generalizability · Tutorial 942 of 972

Transport from Aggregate Target Covariate Margins

See how calibration weights use aggregate target covariate margins—and identify when missing joint information prevents a unique transported effect.

Advanced 12 min read

What You'll Learn

  • Define calibration weights that match published target-population covariate margins.
  • State the causal and sampling conditions needed to transport conditional effects.
  • Calculate a weighted target effect from a single known target margin.
  • Show how different joint distributions can share the same margins but imply different effects.
  • Identify effect structures for which separate margins are sufficient.
  • Recognize calibration feasibility, overlap, and weight-instability limitations.

Why Aggregate Margins Require Care

A target agency may publish the proportion of people in each age group and the proportion with a particular comorbidity, while providing no individual-level target records. Source-study outcomes and treatment effects may be available, but those source participants need not have the same covariate distribution as the target population. Calibration offers a way to reweight source observations so selected covariate averages agree with the published target margins.

Matching margins is not the same as recovering the target population's full covariate distribution. When the causal effect depends on combinations of covariates, separate one-variable margins may leave important joint frequencies unknown. The central question is therefore not only whether weights can match the published numbers, but whether those numbers contain enough information to identify the target effect.

Definition: Let \(S=1\) indicate membership in the source study and \(S=0\) membership in the target population. Let \(X\) denote measured covariates, and let \(Y^a\) be the potential outcome under treatment level \(a\). The target average treatment effect is \(\tau_T=E(Y^1-Y^0\mid S=0)\). Calibration to aggregate target margins chooses source-observation weights so that weighted averages of specified functions of \(X\) equal their known target-population averages. The matched functions may describe separate covariates without determining their joint distribution.

Conditions for a Causal Transport Calculation

Calibration adjusts the covariate distribution represented by the source data; it does not by itself make source treatment comparisons causal or guarantee that effects carry over to the target. A transport calculation needs the relevant causal and sampling conditions. In a randomized source study, treatment assignment identifies source treatment effects under the trial's randomization and consistency conditions. In an observational source study, an appropriate treatment-exchangeability condition and treatment positivity are also needed to identify those effects.

For a covariate set \(X\) sufficient to account for differences in treatment effects between source and target, a standard transport condition is conditional equality of potential-outcome means: \(E(Y^a\mid X,S=1)=E(Y^a\mid X,S=0)\) for each treatment level \(a\), in covariate strata represented in both populations. This condition says that, after conditioning on \(X\), the mean potential outcomes do not differ by study membership. It is not implied by matching covariate margins.

The source must also cover the target covariate values needed for the calculation: the target cannot place positive probability on a covariate pattern that has no source representation. This is transport positivity. For a useful weighting calculation, the calibration constraints must be feasible using the source data, and the resulting weights should be sufficiently stable. These requirements are distinct: weights may match the published margins exactly but still be highly variable or fail to represent the target's relevant covariate patterns.

Identification result: If the source causal effects are identified, conditional mean transportability holds given \(X\), and the target distribution of \(X\) is known on the supported region, then \(\tau_T=E_T\{E(Y^1-Y^0\mid X,S=1)\}\). If only selected marginal averages of \(X\) are known, this expression is identified only when those averages suffice to calculate the expectation of the conditional effect function, or when additional information determines the missing joint distribution.

How Calibration Uses Published Margins

Suppose there are \(n\) source observations, and \(h(X_i)\) is a vector of covariate functions whose target averages have been published. Write those target averages as \(g_T=E\{h(X)\mid S=0\}\). Calibration selects weights \(w_i\) satisfying \(\sum_{i=1}^n w_i h(X_i)/\sum_{i=1}^n w_i=g_T\). For instance, a component of \(h(X_i)\) could be an indicator that person \(i\) is in a specified age group; its target average is the target proportion in that group.

If source conditional treatment effects are estimated by \(\hat\delta(X_i)\), a normalized weighted estimator is \(\hat\tau_T=\sum_{i=1}^n w_i\hat\delta(X_i)/\sum_{i=1}^n w_i\). A calibration procedure also needs a rule for choosing among possible weights, such as minimizing a distance from initial weights subject to the margin constraints. Different rules can produce different weights when the constraints do not uniquely determine them. Exact balance on the supplied functions does not guarantee balance on every other function of \(X\).

Worked Example: Calibrating to One Target Margin

Hypothetical setting. A randomized source study contains 60% of participants with \(X=0\) and 40% with \(X=1\). The published target margin says 30% of the target population has \(X=0\) and 70% has \(X=1\). Suppose the source study identifies conditional mean treatment effects of 2 and 5 outcome units for \(X=0\) and \(X=1\), respectively, and suppose these conditional effects transport to the target.

The source-to-target calibration factors for the two strata are target proportion divided by source proportion: \(w(0)=0.30/0.60=0.50\) and \(w(1)=0.70/0.40=1.75\). These factors convert the source shares into the target shares: \(0.60(0.50)=0.30\) and \(0.40(1.75)=0.70\). The weighted proportions sum to 1.

The transported effect is therefore \(\hat\tau_T=0.30(2)+0.70(5)=0.60+3.50=4.10\) outcome units. Equivalently, using the source shares and calibration factors, the numerator is \(0.60(0.50)(2)+0.40(1.75)(5)=0.60+3.50=4.10\), and the weighted denominator is \(0.30+0.70=1\).

This calculation uses the single binary covariate's full distribution: its two category probabilities are determined by the published proportion for \(X=1\) and the fact that probabilities sum to 1. The causal interpretation additionally depends on the stated identification and transport assumptions, not just on the arithmetic of the weights.

Why Separate Margins May Not Identify the Effect

With two binary covariates, knowing each marginal proportion generally does not determine the four joint cell probabilities. Let \(X_1\) and \(X_2\) be binary, with target marginal probabilities \(P_T(X_1=1)=a\) and \(P_T(X_2=1)=b\). Let \(t=P_T(X_1=1,X_2=1)\). The four cell probabilities must then be \(P_T(1,1)=t\), \(P_T(1,0)=a-t\), \(P_T(0,1)=b-t\), and \(P_T(0,0)=1-a-b+t\).

All four probabilities must be nonnegative. Thus \(t\) must satisfy \(\max(0,a+b-1)\leq t\leq\min(a,b)\). When this interval has more than one value, the marginal information is compatible with multiple joint distributions. If the conditional effect differs across joint cells, its target average can differ across those distributions.

Worked Example: Same Margins, Different Target Effects

Hypothetical setting. The target margins are \(P_T(X_1=1)=0.60\) and \(P_T(X_2=1)=0.50\). Suppose the source data identify conditional effects of 2, 4, 6, and 10 outcome units in the \((X_1,X_2)=(0,0),(0,1),(1,0),(1,1)\) cells, respectively, and those cell-specific effects transport to the target. Assume there is source support for all the target cells under consideration.

Write \(t=P_T(X_1=1,X_2=1)\). The target cell probabilities, in the stated order, are \(t-0.10\), \(0.50-t\), \(0.60-t\), and \(t\). Nonnegativity requires \(t\geq0.10\), \(t\leq0.50\), and \(t\leq0.60\), so the feasible range is \(0.10\leq t\leq0.50\).

For any such joint distribution, the target average effect is \(\tau_T(t)=2(t-0.10)+4(0.50-t)+6(0.60-t)+10t\). Expanding gives \(\tau_T(t)=5.40+2t\) outcome units. At \(t=0.10\), the cell probabilities are \(0,\ 0.40,\ 0.50,\ 0.10\), and the effect is \(5.60\) units. At \(t=0.50\), they are \(0.40,\ 0,\ 0.10,\ 0.50\), and the effect is \(6.40\) units.

Both joint distributions have the same published margins: in each case \(P_T(X_1=1)=0.60\) and \(P_T(X_2=1)=0.50\). Yet they imply different target effects. The margins alone therefore do not identify a unique effect in this example. A calibration algorithm may return a particular joint distribution represented by its weights, but selecting that distribution through an algorithm is not evidence that it is the actual target joint distribution.

When Marginal Information Is Enough

Missing joint information does not always prevent calculation. If the conditional effect function has a structure that makes its expectation depend only on the known margins, the target effect can be identified from those margins. A simple sufficient case is an additive conditional effect with known or identified coefficients and no interaction between the covariates.

Worked Example: An Additive Effect Function

Hypothetical setting. Suppose source data identify the transported conditional effect function \(\delta(x_1,x_2)=2+3x_1+4x_2\) outcome units, and the target margins are again \(E_T(X_1)=0.60\) and \(E_T(X_2)=0.50\). Assume conditional transportability and source support hold.

Taking the target expectation and using linearity gives \(\tau_T=E_T\{\delta(X_1,X_2)\}=2+3E_T(X_1)+4E_T(X_2)\). Substitution yields \(\tau_T=2+3(0.60)+4(0.50)=2+1.80+2.00=5.80\) outcome units.

No joint probability appears in this calculation, so the unknown association between \(X_1\) and \(X_2\) does not affect the answer under this specified additive effect model. This is a substantive restriction on the effect function, not a consequence of calibration. If the effect includes an interaction term, such as \(\gamma X_1X_2\) with \(\gamma\ne0\), its target expectation includes \(\gamma P_T(X_1=1,X_2=1)\), which is not determined by the two margins alone.

A Short Justification for the Sufficiency Claim

The additive example follows directly from linearity of expectation. For a function \(\delta(X_1,X_2)=c+\beta_1X_1+\beta_2X_2\), \[ E_T\{\delta(X_1,X_2)\}=c+\beta_1E_T(X_1)+\beta_2E_T(X_2). \] Thus, if the intercept and coefficients are identified and the two target means are known, every quantity on the right is known. The joint distribution is not needed for this particular functional form.

In contrast, an interaction contribution \(\gamma X_1X_2\) has expectation \(\gamma E_T(X_1X_2)=\gamma P_T(X_1=1,X_2=1)\). Separate margins provide \(E_T(X_1)\) and \(E_T(X_2)\), but generally not \(E_T(X_1X_2)\). This algebra explains why matching both marginal proportions does not automatically identify the target average when effects vary across joint covariate patterns.

Practical Limits and Common Mistakes

Calibration is most defensible when the published margins are accurately defined for the intended target population, align with the source covariates, and include the covariate functions needed for the effect calculation. If target margins are estimated rather than known, their uncertainty should be reflected in uncertainty assessment; treating noisy published estimates as exact can make the final uncertainty appear too small. Margins from different target populations, time periods, or covariate definitions cannot be combined without a justified harmonization.

A frequent mistake is to say, “The weights match age and comorbidity, so they reproduce the target population.” They reproduce the supplied age and comorbidity margins, not necessarily the age-by-comorbidity distribution or other unreported features. Another is to treat a convenient calibration solution as if it were observed target data. The correction is to state precisely which margins are matched, identify which joint or higher-order information remains unknown, and assess whether the conditional effect structure makes that missing information relevant.

  • Check feasibility: the source observations must permit weights satisfying all stated constraints. Incompatible or overly detailed constraints may have no solution.
  • Check support: matching overall margins cannot create source information for target covariate patterns absent from the source.
  • Check weight behavior: extreme weights can make the estimate unstable; some calibration methods can even produce negative weights, which are not ordinary population reweighting probabilities.
  • Check the effect function: if the target average depends on an unreported joint probability, the margins alone do not determine it unless an additional defensible restriction or data source supplies that information.
Key takeaway: Aggregate target margins can calibrate source data to known target averages and may be enough when the transported effect function depends only on those averages. They do not, by themselves, reveal the target joint covariate distribution. A valid transported estimate requires both causal transport conditions and enough target distributional information for the particular effect being averaged.

Check Your Understanding

Answer each question by distinguishing what calibration enforces from what the causal transport calculation requires.

  1. A source has 25% of participants in an age category, while the target margin is 40%. What source-to-target factor would calibrate that category, and what other information is needed to compute a target effect?
  2. For two binary covariates with target margins \(a=0.70\) and \(b=0.60\), state the feasible range for \(t=P_T(X_1=1,X_2=1)\). Does knowing the margins specify \(t\) uniquely?
  3. Why can a conditional effect function \(c+\beta_1X_1+\beta_2X_2\) be averaged using the two covariate means, while an effect containing \(\gamma X_1X_2\) may require joint information?
  4. Explain why exact calibration balance on published margins does not establish conditional transportability.
  5. Name two distinct problems that can arise even when a numerical calibration solution matches every supplied margin.
Need help applying this method to your own data?

Our team supports biostatistics, statistical programming and pharmacometric analyses. Send us a request and we will reply by email.

Request an analysis
Clinical Biostats