Technical Summary: Confounder Detection via Treatment Intent (CDTI)

Problem Statement

In causal inference, Randomized Controlled Trials (RCTs) are the gold standard but are often constrained by cost, time, and ethical limitations, necessitating the use of observational data. However, observational studies face the critical challenge of unobserved confounding (UC): the presence of variables that influence both treatment allocation ($X$) and the outcome ($Y$) but are not recorded in the dataset. Standard adjustment methods (e.g., back-door criterion) fail when such variables are missing.

While domain experts possess knowledge of these unobserved factors (e.g., a physician's perception of a patient's difficulty breathing), they are rarely trained to enumerate them abstractly. Existing methods for eliciting UCs are limited. This paper addresses the gap of systematically identifying candidate unobserved confounders by leveraging the decision-making process of the agent assigning the treatment.

Methodology: Confounder Detection via Treatment Intent (CDTI)

The authors propose a new observational study design called Confounder Detection via Treatment Intent (CDTI). The core premise is that any unobserved confounder $U$ affecting the treatment decision $X$ must be available to the Decision Maker (DM) at the time of the decision. Instead of asking the DM to list confounders in the abstract, the method queries the DM to compare specific pairs of units (patients) to elicit the reasons for differential treatment decisions.

Framework Components

The framework consists of two main strategies:

1. Matching Strategy ($\mathcal{M}$): A principled method for selecting pairs of units $(i, j)$ where $X_i=1$ (treated) and $X_j=0$ (untreated) to present to the expert.
2. Extraction Strategy ($\varepsilon$): A method for processing the expert's reasoning to identify candidate unobserved variables.

Matching Strategies

The paper formalizes three specific matching strategies designed to maximize the probability that the DM's explanation relies on unobserved variables ($U$) rather than observed covariates ($Z$):

• $Z$-matching ($\mathcal{M}_{Z\text{-match}}$): Selects pairs with identical observed covariates ($Z_i = Z_j$). If the treatment differs, the explanation must lie in $U$.
• $\pi$-matching ($\mathcal{M}_{\pi\text{-match}}$): Selects pairs with identical propensity scores ($\pi(Z_i) = \pi(Z_j)$). This allows for matching in high-dimensional spaces where exact $Z$-matching is infeasible.
• $Z$-dominance ($\mathcal{M}_{Z\text{-dom}}$): Selects pairs where the treated unit is strictly "smaller" (less severe) than the untreated unit across all observed dimensions ($Z_i \le Z_j$). Intuitively, if a "healthier" patient receives treatment while a "sicker" one does not, the decision must be driven by unobserved severity markers.

Theoretical Foundations

The authors establish stochastic dominance results (Theorem 1) under specific assumptions (monotonicity of treatment probability in $U$, and log-supermodularity of the joint distribution of $U$ given $Z$). These theorems prove that under the proposed matching strategies, the distribution of unobserved confounders $U$ for the treated unit is stochastically larger than for the untreated unit. This creates a systematic shift that makes $U$ the most likely explanation for the treatment difference.

Furthermore, the paper analyzes Extraction Strategies, defining the probability of "success" (accurately identifying a genuine UC). Theoretical results (Theorems 2 and 3) demonstrate that:

• $Z$-dominance provably increases accuracy over $Z$-matching under perfect extraction.
• $\pi$-matching dominates marginal (random) matching when the observed covariates $Z$ explain a significant portion of the variation compared to $U$.

Key Contributions

1. Formalization of CDTI: Introduction of a new study design that systematically elicits unobserved confounders by querying decision-makers on matched pairs.
2. Theoretical Guarantees: Development of a framework comprising matching and extraction strategies, supported by stochastic-dominance theorems characterizing when these strategies yield informative pairs.
3. Empirical Validation in ICU:
   • Demonstration that Electronic Health Record (EHR) estimates for mechanical ventilation are strongly confounded (estimates suggest increased mortality, contradicting clinical intuition).
   • Validation of the methodology using a semi-synthetic dataset derived from MIMIC-III, where ground truth for unobserved confounders is known (extracted from clinical text notes).
   • Application to real MIMIC-III data to discover clinically plausible confounders (e.g., pleural effusion, dyspnea, pneumonia) that are often missing from structured EHR fields.

Experimental Results

• Synthetic Verification: Experiments confirmed that $Z$-dominance outperforms $Z$-matching when theoretical conditions hold, and that $\pi$-matching outperforms marginal matching when the signal-to-noise ratio of observed covariates is high.
• Semi-Synthetic MIMIC-III: Using clinical notes as a proxy for $U$, the study showed that pairs selected via $\pi$-matching and $Z$-dominance yielded significantly higher success rates in identifying true unobserved confounders compared to random pairing.
• Real-World Application:
  • The method successfully identified 20 high-frequency candidate confounders from clinical notes, 17 of which were deemed clinically plausible (e.g., pulmonary impairment, SIRS, hemorrhage).
  • Effect on ETT: When the detected confounders were added to the adjustment set for estimating the Effect of Treatment on the Treated (ETT) of mechanical ventilation, the estimates changed slightly but did not shift to zero or become statistically significant. The authors attribute this to the noisy and partial nature of text data (concepts are documented for only a minority of patients), suggesting that while the method identifies candidates, the current proxy (binary indicators from notes) may not fully recover back-door validity.

Significance and Claims

The paper claims to provide a theoretical basis for a new observational study design that bridges the gap between raw observational data and domain expertise. Its primary significance lies in:

• Offering a systematic, principled alternative to ad-hoc expert elicitation for uncovering unobserved confounders.
• Providing stochastic dominance guarantees that justify why specific matching strategies (like $Z$-dominance and $\pi$-matching) are more likely to reveal unobserved factors than others.
• Demonstrating a proof of concept in a high-stakes medical setting (ICU), showing that clinical text notes can serve as a proxy for unobserved knowledge, though the authors remain modest about the immediate ability of this method to fully correct causal estimates in the presence of noisy proxies.

The authors explicitly state that the framework detects candidates for unobserved confounders with probabilistic guarantees but does not certify that any specific detected variable is a true confounder or that adding them automatically recovers back-door validity. These remain open questions for future work.