Abstract
Estimating causal effects from observational data is a fundamental task in science, yet it is often hampered by confounding and the inability to distinguish between Markov equivalent causal structures. We present FAS (Find Adjustment Set), a Bayesian method that leverages limited experimental data to identify the correct adjustment set from observational data. By using experimental results as a ground truth to score potential adjustment sets, FAS can resolve ambiguities in causal structure and account for selection bias in experimental populations.
Introduction: The Challenge of Markov Equivalence
In causal inference, the goal is to determine the effect of a treatment $X$ on an outcome $Y$. While observational data is abundant, it is frequently "confounded" by variables that influence both treatment and outcome. To obtain an unbiased estimate, researchers must identify and control for a valid adjustment set.
The primary obstacle is that observational data often supports multiple causal graphs that are "Markov equivalent"—they imply the same set of conditional independencies but different causal effects. Consequently, data alone cannot determine which variables should be included in the adjustment set.
Methodology: The FAS Framework
The proposed FAS algorithm addresses this by incorporating experimental data, such as Randomized Controlled Trials (RCTs). Although RCTs are often too small for precise estimation or are conducted on biased sub-populations, they provide a "gold standard" for the direction and magnitude of causal effects.
Scoring Adjustment Sets
The core of FAS is a Bayesian scoring mechanism. For each candidate adjustment set $Z$, the algorithm calculates the posterior probability that $Z$ is the correct set given both observational and experimental data. The logic is straightforward: if $Z$ is the true adjustment set, the causal effect estimated from observational data using $Z$ should be consistent with the experimental observations.
Handling Selection Bias
Experimental data often suffers from selection bias, where the trial population (e.g., excluding certain age groups) differs from the general population. To address this, the authors introduced SelectionBN. This extension uses a Bayesian Network to model the selection process, allowing the algorithm to "re-weight" observational estimates to match the experimental population's characteristics for a fair comparison.
Experimental Evaluation
The researchers validated FAS using synthetic simulations and real-world medical data.
Synthetic Studies
Using random Semi-Markov Causal Models (SMCMs), the authors tested FAS against scenarios involving no bias, observed selection bias, and latent selection bias. FAS consistently outperformed baseline methods like the "disjunctive set criterion" (VWS), particularly in the presence of selection bias.
Real-World Application: Statins and Myalgia
The method was applied to Electronic Health Record (EHR) data for 100,000 patients and results from the STOMP trial, which studied the effect of statins on muscle pain (myalgia). Without prior medical knowledge, FAS identified Age as the most probable adjustment variable. This aligns with clinical expertise, as age is a known confounder in statin studies, demonstrating the algorithm's ability to recover meaningful causal structures from raw data.
Conclusion and Future Directions
FAS provides a robust framework for causal discovery by reconciling large-scale observational data with limited experimental evidence. It successfully breaks Markov equivalence and handles selection bias, even when key variables are missing from the experimental report.
Future work will focus on scaling the algorithm to high-dimensional datasets, as the search space for adjustment sets grows exponentially. Additionally, extending the framework to handle continuous and mixed data types will enhance its utility in diverse scientific domains.
Summary updated