PCL: Proximal Causal Learning

Andrew YingOpenalex Unknown 2021 · DOI · OpenAlex

Abstract

We fit causal models using proxies. We implement a two-stage proximal least squares estimator. This work builds on the framework established by E.J. Tchetgen Tchetgen, A. Ying, Y. Cui, X. Shi, and W. Miao in "An Introduction to Proximal Causal Learning" (arXiv:2009.10982).

Introduction

In observational studies, the "no unmeasured confounding" assumption is the standard requirement for causal inference. However, in practice, researchers often face hidden factors—such as genetic predisposition or motivation—that influence both treatment and outcome but remain unmeasured. While Instrumental Variable (IV) methods offer a solution, finding valid instruments is notoriously difficult.

The PCL R package provides a practical implementation of Proximal Causal Learning, a framework designed to estimate causal effects in the presence of unmeasured confounders. By leveraging "proxies"—noisy measurements related to hidden variables—the package allows for valid causal inference without requiring the confounders themselves to be directly observed.

Implementation and Methodology

The core of the PCL package is a two-stage proximal least squares estimator. This approach relies on categorizing available proxy variables into two distinct groups:

  1. Treatment-inducing proxies (trt_pxy): Variables related to the unmeasured confounder that primarily influence treatment assignment.
  2. Outcome-inducing proxies (out_pxy): Variables related to the unmeasured confounder that primarily influence the final outcome.

By utilizing these two sets of proxies, the algorithm constructs a "bridge" (formally a Fredholm integral equation of the first kind) to recover the average causal effect.

The pcl() constructor function organizes data into a specialized object by defining outcomes, treatments, and proxy types.

The package workflow involves two primary functions:

  • pcl(): A constructor that organizes the outcome, treatment, proxies, and standard covariates into a specialized object.
  • pclfit(): The estimation engine that fits the model, defaulting to the Proximal Outcome Regression (POR) method.
The pclfit() function takes the PCL object and applies the specified estimation method.

Synthetic Data and Insights

The package documentation demonstrates its utility through synthetic data generated using standard R distributions. These examples highlight a key insight: researchers do not need perfect measurements of hidden variables; having two distinct "views" (the two proxy types) allows the mathematical framework to cancel out the bias.

However, the method requires specific structural conditions: the treatment-inducing proxy should not directly affect the outcome except through the confounder. The PCL package abstracts these complex integral equations into a simple two-step workflow, making advanced causal machinery accessible to practitioners.

Limitations and Future Directions

While powerful, the current version (1.0) of the PCL package has several limitations:

  • Methodological Focus: It primarily supports the POR method, though the broader framework includes Proximal Weighting (PWS) and doubly robust estimators.
  • Binary Treatment: The implementation is currently limited to binary treatment variables, restricting its use for continuous exposures.
  • Proxy Selection: Users must manually categorize proxies; the package does not yet offer automated discovery or validation of independence assumptions.
  • Uncertainty Quantification: The current manual does not detail built-in methods for calculating standard errors or confidence intervals, which are essential for scientific reporting.

Future iterations may expand to include more flexible treatment types and robust uncertainty estimation via bootstrapping.

Summary updated

Abstract

We fit causal models using proxies. We implement two stage proximal least squares estimator. E.J. Tchetgen Tchetgen, A. Ying, Y. Cui, X. Shi, and W. Miao. (2020). An Introduction to Proximal Causal Learning. arXiv e-prints, arXiv-2009 &lt;<a href="https://doi.org/10.48550/arXiv.2009.10982" target="_top">doi:10.48550/arXiv.2009.10982</a>&gt;.
BibTeX
@inproceedings{ying2021pcl,
  author = {Andrew Ying},
  title = {{PCL: Proximal Causal Learning}},
  booktitle = {Openalex Unknown},
  year = {2021},
  doi = {10.32614/cran.package.pcl},
}
Figure