Technical Summary: Coupling Generative Modeling and an Autoencoder with the Causal Bridge

Problem Statement

The paper addresses the challenge of inferring the causal effect of a treatment ($X$) on an outcome ($Y$) in the presence of an unobserved confounder ($U$) that influences both. Traditional causal inference methods often rely on strong assumptions, such as the absence of unobserved confounders or the availability of valid instrumental variables, which may not hold in practice. While recent approaches utilize proxy variables (negative controls)—specifically a treatment control ($Z$) and an outcome control ($W$)—to mitigate this bias, existing methods for estimating the "causal bridge" function (which links proxies to the causal effect) face limitations in theoretical understanding and practical implementation, particularly regarding the efficiency of learning and the handling of complex, nonlinear relationships.

Methodology

The authors propose a framework that couples the causal bridge function with generative modeling and an autoencoder architecture to improve the estimation of treatment effects.

1. Theoretical Refinement of the Causal Bridge

The paper re-examines the assumptions required for the causal bridge function $b(W, x)$ to satisfy the Fredholm integral equation:
$$E(Y |x, z) = E(b(W, x)|x, z)$$
The authors demonstrate that the bridge function does not need to match the underlying structural function pointwise but rather in expectation. They derive a new information-theoretic bound on the average error ($\eta$) of the treatment effect estimate. This bound relates the error to the conditional mutual information $I(U; Z|W, x)$, suggesting that the quality of the bridge approximation depends critically on $W$ being a low-noise representation of the latent confounder $U$.

2. Generative Modeling for Sampling

Instead of relying solely on estimating conditional expectations (as in previous deep proxy causal learning methods), the authors introduce a generative approach. They model the conditional distribution $p(W|x, z)$ using a generative model (e.g., a Conditional GAN or a diffusion model) to sample outcome proxies $W$ given treatment $X$ and treatment proxy $Z$. This allows for the simulation of samples from the latent confounder distribution $p(U|W, x)$ via a function $h(W, x, \epsilon)$, where $\epsilon$ is noise. The bridge function is then formulated as an expectation over these sampled latent variables:
$$b(W, x) \approx E_{\epsilon}[g_Y(x, W, h(W, x, \epsilon))]$$

3. Autoencoder Architecture for Statistical Strength

To enhance the quality of the learned latent representation $h(W, x, \epsilon)$, the authors propose an autoencoder architecture that shares statistical strength across observed variables. The model jointly minimizes a composite loss function:

• Outcome Loss ($L_{\theta_Y}$): Minimizes the error between observed outcomes $Y$ and the bridge prediction.
• Treatment Control Loss ($L_{\theta_X}$): Reconstructs the treatment $X$ from the latent representation.
• Proxy Loss ($L_{\theta_Z}$): Reconstructs the treatment proxy $Z$ from the latent representation.

By sharing the encoder $h_{\theta_U}$ across these tasks, the model leverages information from $X$ and $Z$ to better infer the latent confounder $U$, leading to more robust causal estimates, particularly when sample sizes are small.

4. Extension to Survival Outcomes

The framework is extended to handle survival data (time-to-event outcomes with censoring). Instead of modeling the expected value of $Y$, the method models the hazard function $\lambda(Y | W, x) = \lambda_0(t) \exp(b(W, x))$. The causal effect is estimated via the hazard ratio (HR), optimized using a partial likelihood loss function similar to the Cox proportional hazards model, without requiring a parametric baseline hazard.

Key Contributions

1. Theoretical Bound: The paper provides a new bound on the average error of the causal bridge, linking estimation error to the conditional mutual information between the unobserved confounder and the proxies. It clarifies that the bridge function need not be independent of $Z$ pointwise, but must satisfy an expectation-matching condition.
2. Generative Learning Framework: A novel learning formulation that utilizes generative models to sample from the conditional distribution of proxies, moving beyond the estimation of conditional expectations used in prior state-of-the-art methods (like Deep Feature Proxy Variables).
3. Autoencoder Integration: The introduction of an autoencoder architecture that shares the latent encoder across treatment, proxy, and outcome models, improving the estimation of the causal bridge by leveraging shared statistical strength.
4. Survival Analysis Extension: The first application of the causal bridge framework to survival outcomes, enabling causal inference for time-to-event data in the presence of unobserved confounding.

Experimental Results

The proposed approach was validated on both synthetic and real-world datasets:

• Synthetic Data (Demand and dSprite): On synthetic datasets where ground-truth causal effects are known, the proposed method (CB + AE) significantly outperformed the state-of-the-art Deep Feature Proxy Variable (DFPV) method. The results showed that sampling from $p(W|x, z)$ and utilizing the shared autoencoder reduced the Mean Squared Error (MSE) of the causal effect estimates.
• Real-World Data (Framingham Heart Study): The method was applied to estimate the causal effect of statin use on cardiovascular disease (CVD) events.
  • Standard observational methods (CoxPH with various weighting schemes) produced biased estimates (Hazard Ratio > 1), likely due to residual confounding (high-risk patients being more likely to receive treatment).
  • The proposed Causal Bridge methods (CB and CB + AE) produced Hazard Ratios < 1, consistent with the direction of effect found in a separate Randomized Control Trial (RCT) used as a reference.
  • The CB + AE variant provided tighter confidence intervals and better alignment with the RCT results compared to the CB-only variant, demonstrating the benefit of the autoencoder component.

Significance and Claims

The authors claim that their work advances the theoretical understanding of the causal bridge by relaxing strict independence assumptions and providing a concrete error bound. Practically, the integration of generative models and autoencoders offers a more flexible and robust mechanism for causal inference with proxy variables, particularly in scenarios with limited data or complex nonlinear confounding. The extension to survival analysis broadens the applicability of proximal causal inference to critical biomedical domains. The authors note that while the approach shows superior performance compared to existing proxy-based methods, verifying the underlying assumptions remains a challenge, and the partitioning of covariates into proxies ($W$ and $Z$) requires careful consideration in practical applications.