When Reconstruction Is Not Enough: Doubly Robust Counterfactual and Interventional Mediation Analysis with Learned Multimodal Representations

Introduction

High‑dimensional multimodal health records are increasingly compressed into latent representations for downstream causal analysis. Most existing pipelines train these embeddings to reconstruct the original data, assuming that a good reconstruction automatically preserves the causal structure needed for counterfactual and mediation inference. Cong Cao’s recent work challenges this assumption, demonstrating that reconstruction fidelity offers no guarantee of causal identifiability and can even hide severe bias. The paper develops a theoretical framework for doubly robust counterfactual and interventional mediation analysis, quantifies the failure modes of reconstruction‑only training, and presents a practical workflow that combines modality‑specific variational autoencoders, cross‑fitted estimators, and diagnostic checks.

Why Reconstruction Is Insufficient

The core insight is that causal identification hinges on preserving conditional independence relationships among treatments, outcomes, and confounders. A representation that merely minimizes reconstruction loss may discard or distort the very signals that block back‑door paths. The authors prove two key results:

1. No covariate‑only objective—whether reconstruction, contrastive, or physics‑informed—can certify that a learned representation retains all necessary adjustment information.
2. In a linear structural model, a representation can lose an arbitrarily small amount of reconstruction accuracy while inducing arbitrarily large bias in estimated causal effects.

These findings imply that standard validation metrics (e.g., reconstruction loss) are blind to the identification property required for valid causal inference.

Doubly Robust Estimation with Representation Bias

The paper extends the doubly robust (DR) estimator to the setting where the covariates are replaced by a learned representation \(\phi(X)\). By cross‑fitting nuisance models (propensity scores and outcome regressions) on separate folds, the estimator retains the usual DR property provided the representation is sufficient. The authors decompose the estimator’s error into:

• First‑order representation bias – a term that persists regardless of double robustness or cross‑fitting; it reflects the mismatch between the true confounding structure and what \(\phi\) captures.
• Second‑order remainder – driven by estimation error in the nuisance functions; this term shrinks with sample size.

When \(\phi\) is insufficient, the first‑order bias dominates, causing coverage of confidence intervals to worsen as the sample grows—a counter‑intuitive but critical warning for large‑scale health studies.

Practical Workflow and Diagnostics

To mitigate these risks, the author proposes a step‑by‑step pipeline:

1. Modality‑specific VAEs trained on an external dataset (outside the causal estimation sample) to avoid leakage.
2. Locked phenotypes – pre‑defined clinical sub‑groups that remain fixed during analysis, ensuring interpretability.
3. Cross‑fitted doubly robust estimation using the learned representation as the adjustment set.
4. Select‑then‑confirm pathway analysis – first screen potential mediators, then confirm effects with DR estimators.
5. Portable representations – store the encoder separately so it can be reused across studies without re‑training.
6. Diagnostics:

- Representation insufficiency: compare DR estimates using \(\phi\) versus raw covariates; large discrepancies flag potential bias. - Exposure leakage: assess whether treatment information leaks into \(\phi\) by training a simple classifier on the representation. - Sensitivity analysis: perturb the representation or add/remove weak confounders to gauge robustness. In simulated experiments, discarding a rare but moderately strong confounder caused indirect‑effect confidence intervals to achieve only 2 % coverage at \(n=3{,}000\). Reconstruction loss failed to detect the problem, whereas a lightweight predictive diagnostic identified the issue in every replicate.

Closing Thoughts

Cao’s work underscores a fundamental mismatch between predictive objectives (reconstruction) and causal requirements (identification). By formalizing the bias introduced by insufficient representations and offering a concrete, doubly robust workflow with built‑in diagnostics, the paper equips researchers to build more trustworthy digital‑twin style analyses of multimodal health data. The take‑away for practitioners is clear: validate representations for causal adequacy, not just reconstruction quality, before drawing policy‑relevant conclusions.