Semiparametric Efficient Test for Interpretable Distributional Treatment Effects

Abstract: Distributional treatment effects can be invisible to means: a treatment may preserve average outcomes while changing tails, modes, dispersion, or rare-event probabilities. Kernel tests can detect discrepancies between interventional outcome laws, but global tests do not reveal where the laws differ. We propose DR-ME, to our knowledge the first semiparametrically efficient finite-location test for interpretable distributional treatment effects. DR-ME evaluates an interventional kernel witness at learned outcome locations, returning causal-discrepancy coordinates rather than only a global rejection. From observational data, we derive orthogonal doubly robust kernel features whose centered oracle form is the canonical gradient of this finite witness. For fixed locations, we characterize the local testing limit: DR-ME is chi-square calibrated under the null, has noncentral chi-square local power, and uses the covariance whitening that optimizes local signal-to-noise for discrepancies visible through the selected coordinates. This efficient local-power geometry yields a principled location-learning criterion, with sample splitting preserving post-selection validity. Experiments show near-nominal type-I error, competitive power against global doubly robust kernel tests, and interpretable learned locations that localize distributional effects in a semi-synthetic medical-imaging study.

• The paper introduces DR-ME, a semiparametric efficient test that leverages finite-location projections to localize distributional treatment effects.
• It employs orthogonalization and double robustness to mitigate bias in nuisance estimation while ensuring valid inference via sample splitting.
• Empirical results demonstrate accurate type-I error control and superior power in detecting localized treatment effects in high-dimensional outcomes.

Semiparametric Efficient Testing for Interpretable Distributional Treatment Effects

Motivation and Problem Setting

The paper addresses the challenge of identifying and interpreting distributional treatment effects (DTEs) that are not necessarily reflected in mean contrasts. In many scientific and applied settings, interventions may modify the tails, multimodality, or variance of outcome distributions without affecting means—rendering classical mean-based treatment effect estimation non-informative. This is particularly acute with high-dimensional or structured outcomes (e.g., images, sequences, embeddings), where reductions to summary statistics obscure salient causal effects.

To remedy the limitations of global kernel two-sample tests for causal inference—which, despite their sensitivity, lack localization and interpretability—the paper proposes a new approach: an interpretable, semiparametrically efficient finite-location test for DTEs termed Doubly Robust Mean Embedding (DR-ME).

Methodology: DR-ME and Finite-Location Causal Witnesses

DR-ME operates by evaluating the kernel mean embedding witness for interventional distributional differences at adaptively selected, finite sets of outcome locations. For potential outcomes $Y(0), Y(1)$, and a bounded positive definite kernel $k_y$, the mean embeddings $x(a) = \mathbb{E}[\phi_y(Y(a))]$ define the global discrepancy, with the kernel witness function $$w(y) = \mathbb{E}[k_y(y, Y(1))] - \mathbb{E}[k_y(y, Y(0))]$$ capturing pointwise distributional differences.

Instead of a global test, the method projects $w(\cdot)$ onto a small set of locations $V = (v_1, \ldots, v_J)$, yielding interpretable coordinates $\mu_V = (w(v_1), \ldots, w(v_J)) \in \mathbb{R}^J$. These serve not only as test statistics but directly indicate where the treatment effect manifests in outcome space.

Key innovations include:

• Orthogonalization and Double Robustness: The observed-data feature construction uses an augmented inverse-propensity score structure, ensuring robustness to nuisance estimation of both the propensity model and outcome regressions. The canonical gradient is derived as a vector-valued influence function with cross-fitting to further eliminate first-order bias due to nuisance estimation error.
• Semiparametric Efficiency: The canonical gradient's covariance defines the local geometry of the finite-location testing problem, enabling construction of an omnibus Hotelling-type test statistic with chi-square calibration and optimal local power under finite-dimensional alternatives. The method attains semiparametric efficiency for the finite-location parameter of interest.

Statistical Theory: Local Testing Geometry and Location Learning

The theoretical results characterize the DR-ME test's asymptotic distribution and local power under local alternatives. Under standard regularity and identification conditions, for fixed outcome locations, the DR-ME statistic (with sample splitting for post-selection validity) converges to a chi-square distribution under the null. Power against local alternatives is a noncentral chi-square, with noncentrality parameter given by the covariance-whitened local drift.

This local testing geometry elucidates two fundamental points:

1. Covariance Whitening is Essential: Location learning is guided by maximizing the local signal-to-noise ratio, as measured by the covariance-whitened witness magnitude. This penalizes redundant, high-variance, or uninformative directions in outcome space, favoring locations where discrepancies are not only large but statistically distinguishable.
2. Interpretable, Valid Post-Selection Inference: By employing sample splitting for location selection, the chi-square calibration holds conditionally on the selected locations, providing valid type-I error control regardless of the (possibly data-driven) selection mechanism.

Empirical Results

The paper's empirical evaluation demonstrates several critical outcomes:

• Type-I Error Control: On synthetic and semi-synthetic observational data with confounding, the split-sample DR-ME test maintains type-I error near nominal levels, whereas naive procedures without orthogonalization, double robustness, or sample splitting severely over-reject.
• Power and Location Learning: When discrepancies are localized in the outcome space, DR-ME with learned locations significantly outperforms global kernel treatment effect tests and finite-location tests with random or non-whitened location selection. The uniform consistency of the learning criterion is established both theoretically and in simulation.
• Interpretability: In a semi-synthetic medical imaging experiment, DR-ME identifies image locations where rare but clinically relevant treatment-induced effects occur—even when these are invisible to both mean contrasts and global tests. The selected regions correspond to pathological features engineered in the data construction, providing actionable and interpretable localization.

Computational Considerations

The method offers competitive or superior runtime compared to global kernel-based DR tests, particularly at large sample sizes or with low $J$. The final test is a Hotelling statistic in $\mathbb{R}^J$, providing computational advantages for high-dimensional or structured outcomes. Both finite-dictionary and gradient-based implementations are discussed, with empirical benchmarks guiding the choice per application regime.

Implications and Future Directions

The development of DR-ME bridges a methodological gap between global kernel hypothesis testing and interpretable, localizable analysis of complex treatment effects under observational data. By aligning semiparametric efficiency theory with interpretable, finite-dimensional projections, the method delivers both statistical optimality and scientific transparency.

The results have several implications:

• Experimental and Causal Inference: Enables discovery and communication of subtle, non-mean-based causal effects, essential in domains such as biomedical imaging, genomics, or high-throughput screening.
• Scalability and Adaptability: Location learning via covariance whitening and finite-location testing is extensible to dictionary-based or regularized search classes, suitable for structured outcome domains (e.g., sequences, graphs).
• Theoretical Generalization: Future work can extend the approach to longitudinal, missing data, or instrumental variable settings—requiring new canonical gradients and extensions of the local asymptotic analysis. Representation learning (e.g., deep kernels) is also compatible, conditional on decoupling training and test data to preserve inference validity.

Conclusion

This work establishes a rigorous, semiparametrically efficient framework for interpretable, finite-location testing of distributional treatment effects. The DR-ME statistic, supported by orthogonalized estimation and a local testing theory framed by covariance whitening, offers both strong statistical properties and scientific interpretability. The experimental evaluation confirms calibration, power, and localization benefits, with a range of practical implications for causal inference with complex structured outcomes. The methodology paves the way for further advances in interpretable, robust, and scalable distributional causal analysis in high-dimensional and structured domains.