Transporting Trial Evidence Under Posterior Drift and Possible Hidden Confounding

Abstract: Randomized trials provide internally valid treatment-effect evidence, but trial participants may not represent the target population. In contrast, observational studies are often closer to the target population, but their treatment assignment may be affected by possible hidden confounding. We develop a robust posterior-drift framework for estimating the average treatment effect in an observational target population when exact conditional-effect transportability may fail. The framework represents observational conditional potential-outcome regressions as their randomized-trial counterparts plus source-specific drifts. The randomized trial serves as an internally valid anchor, while the observational study supplies the target covariate distribution and partial information about the target causal contrast. To account for possible hidden confounding, we consider a Rosenbaum-type uncertainty set induced by a sensitivity parameter on the generalized propensity score and estimate the drift through a minimax worst-case risk criterion. We derive efficiency results in auxiliary regimes, establish uniform concentration and near-optimality guarantees for the minimax estimator, and handle general parametric and smooth nonparametric drift classes. Simulations and an ACTG 175--WIHS application show that the proposed analysis yields more cautious and interpretable target-population effect estimates than exact-transportability analyses.

• Discover a robust method for estimating treatment effects in observational target populations using randomized controlled trials (RCT) and observational studies (OS).
• The proposed framework incorporates RCT as an internally valid anchor, while addressing posterior drift and hidden confounding through minimized estimation error, particularly ensuring efficiency.
• The findings underscore the practical significance of the comprehensive formula, including a test case applying the method to ACTG 175 and WIHS cohorts.

Motivation and problem

Randomized controlled trials (RCTs) deliver internally valid treatment-effect estimates, but trial participants are frequently unrepresentative of the populations where treatment decisions are made. Observational studies (OSs) are typically closer to such target populations, yet their treatment assignment may be hidden-confounded. Existing transportability methods largely rely on exact conditional-average-treatment-effect (CATE) transportability—equality of the trial and target CATE functions after covariate adjustment—an assumption that is difficult to defend when eligibility criteria, adherence patterns, clinical practice, or latent health characteristics differ across sources. The paper under review addresses the resulting gap: estimating the average treatment effect (ATE) $\tau = \mathbb{E}\{Y(1)-Y(0)\mid G=0\}$ in an observational target population when cross-source posterior drift is unknown and OS assignment may be unmeasured-confounded. The framework treats the RCT as an internally valid anchor for the baseline response surface, uses the OS for the target covariate distribution, and calibrates hidden confounding through a Rosenbaum-type sensitivity analysis embedded in a minimax estimation criterion.

Setup and identification structure

The two-source superpopulation model assumes i.i.d. draws of $(X, A, Y(0), Y(1), G)$, with $G=1$ indexing the RCT. The authors define source-specific propensity scores $e_g(x)$, sampling scores $\pi(x)$, and conditional mean potential outcomes $\mu_a$ (RCT) and $\tilde\mu_a$ (OS). The formulation accommodates both nested and non-nested designs via the interpretation of $\pi(x)$ as a participation or membership probability. Only RCT internal validity is assumed throughout: randomization within the trial plus positivity.

The key modeling device is a posterior-drift assumption. In the baseline linear specification,

$$\tilde\mu_0(X)=\mu_0(X)+\alpha^\top X,\qquad \tilde\mu_1(X)=\mu_1(X)+\beta^\top X,$$

so that the OS CATE satisfies $\tilde\Delta(X)=\Delta(X)+\gamma^\top X$ with $(X, A, Y(0), Y(1), G)$0. This reduces the unknown cross-source discrepancy to a finite-dimensional drift coefficient, while avoiding any claim that drift vanishes.

Auxiliary regimes: known drift and OS unconfoundedness

Two auxiliary regimes build intuition before the main construction. First, when the linear drift $(X, A, Y(0), Y(1), G)$1 is known, the efficient influence functions (EIFs) for $(X, A, Y(0), Y(1), G)$2 are derived under (a) RCT validity plus drift only, and (b) additionally OS unconfoundedness. The second EIF combines trial and observational residual terms through variance-ratio weights $(X, A, Y(0), Y(1), G)$3 and composite weights $(X, A, Y(0), Y(1), G)$4 that mix trial and observational propensity information.

A notable result is an explicit efficiency comparison: incorporating the RCT under the linear-drift structure never hurts relative to OS-only estimation. The bound difference $(X, A, Y(0), Y(1), G)$5 equals a nonnegative expectation involving $(X, A, Y(0), Y(1), G)$6 and analogous terms, so the gain vanishes only in degenerate cases. This formalizes the value of trial anchoring even when $(X, A, Y(0), Y(1), G)$7 is already identified from the OS alone.

Estimation proceeds by source-stratified sample splitting with plug-in nuisance estimators, and consistency holds under sup-norm convergence of nuisances to their true limits. When $(X, A, Y(0), Y(1), G)$8 is unknown but the OS is unconfounded, a plug-in estimator regresses the difference between OS and RCT CATE estimates on $(X, A, Y(0), Y(1), G)$9; its $G=1$0 error is bounded by $G=1$1, where $G=1$2 reflects sub-Gaussian tail rates of the underlying CATE estimators, and consistency of the final ATE estimator follows.

Minimax estimation under hidden confounding

In the main regime, OS confounding makes $G=1$3 unidentified from observed OS regressions. The authors exploit the transformed outcome

$G=1$4

which satisfies $G=1$5 for the generalized propensity score $G=1$6. Because $G=1$7 is unobservable, they restrict it to an odds-ratio neighborhood of the ordinary propensity score $G=1$8, parameterized by a sensitivity level $G=1$9: the reciprocal generalized weight $e_g(x)$0 varies over $e_g(x)$1 with endpoints built from $e_g(x)$2 and $e_g(x)$3.

The drift coefficient is estimated by solving

$e_g(x)$4

and plugged into the M-estimator for $e_g(x)$5. Importantly, because OS confounding prevents consistent recovery of the true $e_g(x)$6, the stated objective is minimax optimality rather than convergence to $e_g(x)$7: the empirical worst-case risk uniformly concentrates around the population supremum loss at rate $e_g(x)$8, and the population sup-risk achieved by $e_g(x)$9 is within twice this rate of the constrained optimum. A supplementary proposition quantifies the additional approximation error from estimating $\pi(x)$0, showing the gap is driven by propensity-score estimation error under positivity and boundedness conditions. A deliberate design choice here is to avoid estimating the OS causal contrast from OS data alone and then applying worst-case optimization; instead the RCT anchors the contrast and only the drift is reconciled robustly.

Extensions beyond linearity

The minimax construction extends to general finite-dimensional parametric drift classes (vector spaces containing nonlinear transformations or interactions), with concentration bounds carrying over essentially unchanged since the proof relies on uniform control arguments independent of linearity; penalized variants stabilize optimization. For nonparametric drifts in a Hölder class $\pi(x)$1, a tensor-product spline sieve yields the excess-risk bound

$\pi(x)$2

with sieve level $\pi(x)$3 of order $\pi(x)$4. The authors note plainly that this requires a user-specified smoothness $\pi(x)$5, and that fully adaptive selection (cross-validation or Lepski-type tuning) within this minimax-transport framework remains open.

Simulation findings

The simulations generate Gaussian covariates with an unobserved confounder $\pi(x)$6 entering both OS treatment assignment and both potential outcomes, with bias.strength controlling linear CATE drift and u.strength controlling confounding severity; trial participation depends on both $\pi(x)$7 and $\pi(x)$8. Across total sample sizes from 600 to 10,000 and six confounding levels, the proposed estimator generally outperforms a no-drift benchmark ($\pi(x)$9), with larger gains in larger samples and appropriately calibrated $\mu_a$0. A characteristic U-shaped calibration pattern emerges in $\mu_a$1: too small a value under-adjusts for confounding, too large introduces conservatism, and intermediate values perform best. Supplementary experiments show the no-splitting implementation dominates the splitting-based one uniformly in these scenarios—sample splitting supports the theory but is conservative in finite samples—and confirm robustness to bounded-support designs, random-forest nuisance estimation, nonlinear nuisances, and quadratic drift. The authors candidly flag that with unbounded Gaussian covariates the population-level interpretation of $\mu_a$2 is delicate, treating those runs primarily as algorithmic comparisons.

Case study: transporting ACTG 175 evidence to WIHS

The application transports evidence from ACTG 175 (women subset, $\mu_a$3; contrast of ddI-containing regimens versus AZT monotherapy; outcome CD4 count at roughly 20 weeks) to the WIHS cohort (429 participants after exclusions, CD4 at approximately six months). Before fitting the drift model, the doubly robust homogeneity test rejects exact CATE transportability for all three ACTG reference samples—women-only: $\mu_a$4, $\mu_a$5; men-only: $\mu_a$6, $\mu_a$7; full sample: $\mu_a$8, $\mu_a$9. The rejection cannot distinguish posterior drift from WIHS confounding, which motivates the joint treatment of both concerns.

Estimated WIHS-target ATEs decline monotonically in $\tilde\mu_a$0 under both nuisance-learning strategies:

Nuisance	Baseline	$\tilde\mu_a$1	$\tilde\mu_a$2	$\tilde\mu_a$3	$\tilde\mu_a$4
Linear	38.33 (16.01)	30.27 (16.56)	22.51 (18.50)	14.70 (20.75)	−30.74 (26.86)
Random forest	36.12 (15.06)	35.76 (15.01)	33.22 (15.20)	30.69 (15.44)	15.01 (17.99)

Three conclusions follow directly. First, drift-adjusted estimates fall below the naive-transportability baselines, corroborating the homogeneity-test rejections. Second, increasing $\tilde\mu_a$5 produces increasingly conservative adjustments, as designed. Third, effects remain positive under moderate sensitivity levels except for the linear-nuisance analysis at very large $\tilde\mu_a$6; given prior clinical evidence favoring active therapy, the negative estimates at $\tilde\mu_a$7 and $\tilde\mu_a$8 are best interpreted as overly conservative scenarios rather than plausible conclusions. A data-guided simulation calibrated to the empirical covariate distribution reproduces the downward pattern in $\tilde\mu_a$9 with MSE-minimizing sensitivity parameters above one, supporting the reading that nontrivial robust adjustment corrects overstatement induced by indiscriminate exact transportability. The central practical implication is that ignoring drift and hidden confounding can overstate transported effects by a clinically meaningful margin—in this application, roughly 8–16 CD4 cells/mm³ depending on nuisance strategy and $\pi(x)$0.

Limitations and open questions

Several limitations are acknowledged within the paper itself. The minimax guarantees concern near-optimality of the worst-case population risk, not consistency for the true drift coefficient $\pi(x)$1 under confounding; the interpretation of results therefore depends on how well $\pi(x)$2 captures the actual degree of hidden confounding, and no data-driven calibration procedure for $\pi(x)$3 is provided. The nonparametric extension requires a prespecified smoothness level and leaves adaptive tuning undeveloped. The Gaussian simulation design complicates the exact population-level meaning of $\pi(x)$4, though bounded-support experiments partially address this. Inference for the final ATE relies on bootstrap standard deviations rather than distributional theory for the full minimax-plus-plug-in procedure, and the case study's rejection of homogeneity cannot disentangle drift from confounding—the very ambiguity the method absorbs rather than resolves. Whether the framework extends cleanly to longitudinal treatments or more than two data sources is left open.

Conclusion

This work formulates target-population ATE estimation under simultaneous unknown posterior drift and possible OS hidden confounding, a setting not covered by standard transportability estimators, external-control borrowing, or single-source robust analyses. Its technical core is a Rosenbaum-type uncertainty set combined with uniform concentration and near-optimality guarantees for a minimax drift estimator anchored by randomized-trial CATE information, extended through parametric classes and spline sieves to smooth nonparametric drifts. Simulations and the ACTG 175–WIHS analysis demonstrate concretely that exact-transportability analyses can overstate target-population effects, while the proposed procedure yields more cautious and interpretable estimates whose sensitivity to $\pi(x)$5 is transparent and clinically assessable.