Proximal Causal Learning with Proxy Variables

This direction is merged into another direction: proximal causal inference. Its last version stays readable and citable here.

Version 1 of 1109 papers170 conceptsfrozen 2026-10-04

In the ideal world of causal inference, we observe every variable that influences both a treatment and its outcome. This "no unmeasured confounding" assumption allows us to use the g-formula to calculate the true effect of an intervention. In the real world, however, we are often haunted by "hidden pipers"—latent variables like "patient frailty," "socioeconomic status," or "unrecorded clinical intuition" that influence both the cause and the effect, biasing our results.

Proximal Causal Learning (PCL) provides a mathematical escape hatch. As established in An Introduction to Proximal Causal Learning (2020), researchers can identify causal relationships even when key confounders are unmeasured, provided they have access to proxies: imperfect, noisy measurements that serve as "views" of the hidden factors.

The Intuition: Two Views of a Hidden Variable

The core breakthrough of proximal inference, championed by Eric Tchetgen Tchetgen and Wang Miao, is the dual-proxy mechanism.

If you only have one proxy for a hidden confounder $U$, you generally cannot identify the causal effect because you can't distinguish the signal of the confounder from the noise of the proxy. However, if you have two independent proxies that "see" $U$ from different angles, you can mathematically cancel out the bias. These are typically categorized as:

  1. Treatment Proxies (Z): Variables related to the hidden confounder and the treatment, but which do not directly affect the outcome (e.g., a "negative control exposure").
  2. Outcome Proxies (W): Variables related to the hidden confounder and the outcome, but which are not directly affected by the treatment (e.g., a "negative control outcome").

By satisfying a "completeness condition" between these two sets, the framework allows us to recover the true causal effect as if the hidden variable $U$ had been measured all along.

The Mechanism: Bridge Functions

How do we actually use these proxies? The heavy lifting is done by bridge functions. As explained in Semiparametric Proximal Causal Inference, a bridge function acts as a computational link that maps the relationship between the observed proxies and the outcome.

Mathematically, this involves solving a Fredholm integral equation of the first kind. This sounds daunting, but the intuition is elegant: we are looking for a function of our observed proxies that, when averaged over the distribution of the hidden confounder, equals the mean outcome we would see if we could actually observe that confounder.

Recent software implementations, such as the PCL R package by Andrew Ying, use two-stage proximal least squares to solve these equations, making the theory accessible for practical data analysis.

Robustness and Flexibility

A major advantage of this direction is its resilience to model errors. Hongxiang Qiu and colleagues have developed doubly robust estimators for proximal learning (Doubly Robust Proximal Synthetic Controls). These estimators remain accurate even if either the outcome bridge function or the treatment weighting model is incorrectly specified, providing a safety net for researchers working with messy observational data.

The framework is also remarkably flexible across different data types:

Beyond the Dual-Proxy Requirement

While the dual-proxy setup is the standard, new research is pushing the boundaries of what counts as a proxy. Liyuan Xu and Arthur Gretton proved in Kernel Single Proxy Control for Deterministic Confounding that if a system is deterministic, the outcome itself can serve as a second proxy, meaning you only need one external proxy variable.

Similarly, Baohong Li and Kun Kuang introduced the Two-Stage Shadow Inclusion (2SSI) method (Two-Stage Shadow Inclusion Estimation), which repurposes the residual of a treatment as a "shadow variable" to correct for both latent confounding and selection bias simultaneously.

By shifting the focus from "measuring everything" to "leveraging what we have," Proximal Causal Learning turns the presence of noisy, auxiliary data from a nuisance into a powerful tool for causal discovery.

Papers in this version

Showing 60 of 109 papers.

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_proximal_causal_learning_with_proxy_vari_d10fc810,
  title = {Proximal Causal Learning with Proxy Variables},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/proximal-causal-learning-with-proxy-variables/rdr_d10fc810d5b549c5b038c4df9737a1b5},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_proximal_causal_learning_with_proxy_vari_d10fc810_v1,
  title = {Proximal Causal Learning with Proxy Variables},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/proximal-causal-learning-with-proxy-variables/rdr_d10fc810d5b549c5b038c4df9737a1b5/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}