Object Conditioning for Causal Inference

David D. Jensen, Javier Burroni, Matthew J. RattiganUAI 2019

Introduction

In the field of causal inference, a persistent challenge is the presence of "latent" confounders—unmeasured variables that influence both the cause ($X$) and the effect ($Y$), leading to spurious correlations. While social scientists have long employed "design-based" methods like within-subjects studies to address this, these techniques have historically been difficult to reconcile with formal causal graphical models (CGMs).

In "Object Conditioning for Causal Inference," David Jensen, Javier Burroni, and Matthew Rattigan formalize a technique called object conditioning. They demonstrate that conditioning on the identity of an object (e.g., a specific person) rather than its measured attributes is a mathematically sound way to eliminate bias from hidden variables without inducing the "collider bias" that often plagues traditional statistical adjustments.

Methodology: Plate Models and Object Identity

The authors utilize Plate Models to represent hierarchical data structures, such as a "User" plate containing multiple "Post" instances. The core innovation is the definition of object conditioning: instead of conditioning on a specific variable value (e.g., User.Skill = High), the researcher conditions on the identity of the object itself (e.g., User = Robert456).

Hierarchical plate models illustrating the relationship between parent objects (P) and child instances (C). (a) shows a measured parent variable, (b) a latent parent variable, and (c) a latent variable within the child plate.

By shifting the focus to object identity, the authors move beyond standard logic of conditional independence to the principle of exchangeability. They prove that conditioning on the parent object makes child observations (e.g., different posts by the same user) exchangeable. Consequently, any remaining association between $X$ and $Y$ within those observations can be reliably attributed to a causal effect rather than hidden traits. The authors emphasize that treating an "ID variable" as a standard random variable is insufficient; the unique semantics of object conditioning are required to apply d-separation correctly to multi-object units.

Key Insights and Results

The paper provides a theoretical "free lunch" for causal inference. Conditioning on an object's identity implicitly adjusts for every stable attribute of that object simultaneously, including those that are unmeasured or even unknown to the researcher.

Crucially, the authors resolve the Collider Paradox. In a standard directed acyclic graph (DAG), conditioning on a "collider" (a variable caused by both $X$ and $Y$) induces a spurious correlation. However, the authors prove that if the collider is a property of a parent object, conditioning on the object identity does not induce this bias.

Three canonical causal structures analyzed in the paper: (a) the latent variable Z is a confounder, (b) Z is a mediator, and (c) Z is a collider.

The paper provides formal proofs for three canonical causal structures shown in Figure 5:

  1. Confounder: Where the latent variable $Z$ influences both $X$ and $Y$.
  2. Mediator: Where $X$ influences $Z$, which in turn influences $Y$.
  3. Collider: Where both $X$ and $Y$ influence $Z$.

In all three cases, object conditioning produces valid causal estimates, provided the latent traits remain constant across the observations of that object.

Conclusion

This work unifies several disparate techniques—including within-subject designs, multi-level models (MLMs), and Difference-in-Differences (DID)—under a single theoretical framework. By formalizing object conditioning, the authors provide a path for extending d-separation rules to hierarchical data, which is essential for the development of automated causal discovery algorithms. While the method assumes that latent traits remain constant over time, it offers a powerful tool for extracting causal insights from complex, real-world datasets like Stack Overflow or longitudinal social studies.

Summary updated

BibTeX
@inproceedings{jensen2019object,
  author = {David D. Jensen and Javier Burroni and Matthew J. Rattigan},
  title = {{Object Conditioning for Causal Inference}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {2019},
}
Figure