Abstract
Proximal causal inference is a recently proposed framework for evaluating causal effects in the presence of unmeasured confounding. For point identification of causal effects, it leverages a pair of so-called treatment and outcome confounding proxy variables, to identify a bridge function that matches the dependence of potential outcomes or treatment variables on the hidden factors to corresponding functions of observed proxies. Unique identification of a causal effect via a bridge function crucially requires that proxies are sufficiently relevant for hidden factors, a requirement that has previously been formalized as a completeness condition. However, completeness is well-known not to be empirically testable, and although a bridge function may be well-defined, lack of completeness, sometimes manifested by availability of a single type of proxy, may severely limit prospects for identification of a bridge function and thus a causal effect; therefore, potentially restricting the application of the proximal causal framework. In this paper, we propose partial identification methods that do not require completeness and obviate the need for identification of a bridge function. That is, we establish that proxies of unobserved confounders can be leveraged to obtain bounds on the causal effect of the treatment on the outcome even if available information does not suffice to identify either a bridge function or a corresponding causal effect of interest. Our bounds are non-smooth functionals of the observed data distribution. As a consequence, in the context of inference, we initially provide a smooth approximation of our bounds. Subsequently, we leverage bootstrap confidence intervals on the approximated bounds. We further establish analogous partial identification results in related settings where identification hinges upon hidden mediators for which proxies are available.
Introduction
In causal inference, the primary objective is to identify the effect of a treatment on an outcome. However, the presence of unobserved confounders—latent variables influencing both treatment and outcome—often biases results. While Proximal Causal Inference (PCI) offers a solution by using "proxies" for these hidden variables, it traditionally relies on a strict "completeness condition." This condition assumes proxies are sufficiently rich to capture all variation in the hidden confounder, a requirement that is notoriously difficult to test empirically. This paper proposes a shift from point identification to partial identification, deriving valid bounds for causal effects even when proxies are sparse or the completeness condition fails.
The Proximal Causal Framework
The standard proximal framework utilizes two types of proxies: an outcome confounding proxy ($W$) and a treatment confounding proxy ($Z$). These proxies are used to solve for a "bridge function" that links observed data to the latent confounding structure.
As shown in Figure 1, the proxy $W$ is influenced by the latent confounder $U$ but does not directly affect the treatment $A$, satisfying certain exclusion restrictions necessary for point identification. However, if these restrictions are violated or if only one type of proxy is available, point identification becomes impossible.
Methods for Partial Identification
To overcome the limitations of point identification, the authors introduce several strategies to calculate causal bounds:
- Single Proxy Bounds: Deriving the Average Treatment Effect (ATE) using only one proxy (either $W$ or $Z$), which is highly valuable when the dual-proxy requirement of standard PCI cannot be met.
- Robustness to Invalid Proxies: Traditional PCI fails if proxies are "invalid" (e.g., they directly influence the outcome or are influenced by the treatment). The proposed method can still yield informative bounds in these scenarios.
- Bridge Function Obviation: By focusing on functionals of the observed data rather than solving complex integral equations for bridge functions, the method simplifies the computational burden and avoids the need for completeness.
- Smooth Inference via LogSumExp: Because the derived bounds involve non-smooth "max" and "min" functions, standard bootstrapping is unreliable. The authors employ the LogSumExp (LSE) function to create a smooth approximation, enabling valid 95% confidence intervals.
Empirical Results
Synthetic Simulations
In controlled environments where point identification was theoretically impossible due to "thin" proxies, the proposed bounds consistently contained the true causal effect. The simulations demonstrated that as sample size increases, confidence intervals narrow around the theoretical bounds, confirming the consistency of the approach.
The SUPPORT Study
The framework was applied to the SUPPORT dataset to evaluate the effect of Right Heart Catheterization (RHC) on 30-day survival for critically ill patients. The analysis found a negative causal effect (ATE bounds between -2.28 and -0.04), suggesting that RHC may actually decrease survival rates. This finding aligns with historical medical studies that led to a re-evaluation of RHC in intensive care settings.
Conclusion
This research provides a robust alternative to point identification in proximal causal inference. By moving to partial identification, researchers can extract meaningful causal information from imperfect data without relying on untestable completeness assumptions. While the current implementation focuses on categorical variables and is sensitive to smoothing parameters, it represents a significant step toward more flexible and reliable causal analysis in the presence of unmeasured confounding.
Summary updated