Instrumental Variable-based Identification for Causal Effects using Covariate Information

Yuta KawakamiAAAI 2021 · OpenReview

Abstract

This paper addresses the challenge of identifying causal effects in randomized controlled trials (RCTs) characterized by noncompliance. Traditionally, point identification in these settings requires restrictive assumptions like monotonicity. We propose a novel identification condition that leverages an instrumental variable (IV) and a proxy variable of potential outcomes. By demonstrating that the joint probabilities of potential outcomes can be uniquely determined through matrix inversion of observed data, this method allows for consistent estimation of causal effects without relying on monotonicity, independence between potential outcomes, or gain equality assumptions.

Introduction

In randomized trials, noncompliance occurs when participants do not adhere to their assigned treatment groups. While Intention-to-Treat (ITT) analysis provides a valid estimate of the effect of assignment, it often fails to capture the true effect of the treatment itself. Instrumental Variable (IV) methods are typically employed to address this, but they generally only provide bounds for the causal effect unless strong assumptions, such as monotonicity (the assumption that no one does the opposite of their assignment), are invoked. However, monotonicity is often violated in real-world social and medical contexts. This research explores whether additional covariate information, specifically a proxy variable, can achieve point identification without such constraints.

The Proxy Variable Framework

The proposed methodology introduces a proxy variable, $W$, which serves as a noisy measurement of the latent factors $U$ that influence both treatment selection ($X$) and the final outcome ($Y$).

Causal diagrams illustrating the proposed framework. Left: The proxy variable W as a descendant of latent confounder U. Right: W as a descendant of the potential outcomes (Yx1, Yx0).

As shown in the causal diagrams above, the framework assumes that the instrumental variable $Z$ affects the outcome $Y$ only through the treatment $X$. The key innovation is the inclusion of $W$, which is conditionally independent of $Z$ given the latent response types. These response types categorize individuals into four groups: those who always get the disease ("Doomed"), those who only get it if treated ("Causative"), those who only get it if untreated ("Preventive"), and those who never get it ("Immune").

Identification via Matrix Inversion

The core technical contribution is Theorem 1, which establishes the conditions for point identification. The author constructs matrices ($P_x$ and $Q_x$) based on the observed joint probabilities of $(Z, X, Y, W)$.

If the proxy variable $W$ has at least four categories (corresponding to the four potential outcome types) and the relationship between $W$ and the latent types is sufficiently strong to ensure matrix invertibility, the joint probabilities of the potential outcomes can be uniquely solved. This approach is particularly powerful because the invertibility of these matrices is a testable condition using observed data, providing researchers with a way to verify the validity of their proxy variable.

Conclusion and Implications

By moving beyond the monotonicity assumption, this work significantly expands the toolkit for causal inference in the presence of noncompliance. The ability to calculate the "probability of necessity and sufficiency" has profound implications for personalized medicine and legal reasoning, where understanding individual-level response types is critical.

While the current framework focuses on binary treatments and outcomes, it provides a robust foundation for future extensions to multi-categorical variables. The primary practical challenge remains the selection of an appropriate proxy variable $W$ that is sufficiently informative of the latent response types without being directly influenced by the instrument $Z$.

Summary updated

Abstract

This paper deals with the identification problem of causal effects in randomized trials with noncompliance. In this problem, generally, causal effects are not identifiable and thus have been evaluated under some strict assumptions, or through the bounds. Different from existing studies, we propose novel identification conditions of joint probabilities of potential outcomes, which allow us to derive a consistent estimator of the causal effect. Regarding the identification conditions of joint probabilities of potential outcomes, the assumptions of monotonicity (Pearl, 2009), independence between potential outcomes (Robins & Richardson, 2011), gain equality (Li & Pearl, 2019) and specific functional relationships between cause and effect (Pearl, 2009) have been utilized. In contrast, without such assumptions, the proposed conditions enable us to evaluate joint probabilities of potential outcomes using an instrumental variable and a proxy variable of potential outcomes. The results of this paper extend the range of solvable identification problems in causal inference.
BibTeX
@inproceedings{kawakami2021instrumental,
  author = {Yuta Kawakami},
  title = {{Instrumental Variable-based Identification for Causal Effects using Covariate Information}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  url = {https://openreview.net/forum?id=hGwC9qws4u},
}
Figure