Causal Inference in Observational Data

Pranjul Yadav, Lisiane Prunelli, Alexander Hoff, Michael Steinbach, Bonnie Westra, Vipin Kumar, Gyorgy SimonarXiv 2016 · arXiv

TL;DR

This paper introduces a novel causal rule mining framework that transitions data mining from simple association to true causal inference within patient subpopulations. By defining "closed intervention sets" and introducing efficient pruning strategies, the framework successfully identifies optimal combinations of medical treatments from observational Electronic Health Record (EHR) data while correcting for systemic biases.

What Problem Does This Paper Address?

As the global population ages, clinicians increasingly treat patients suffering from multiple co-occurring chronic conditions (multimorbidity). Finding the optimal combination of treatments—such as multi-drug cocktails—is a massive combinatorial challenge.

Traditional data mining techniques like association rule mining excel at finding patterns in large datasets, but they only identify correlations, not causation. In healthcare, relying on correlation can be dangerous due to confounding by indication: a drug might appear associated with higher mortality simply because it is routinely prescribed to the sickest patients. This paper addresses the challenge of scaling rigorous causal inference—traditionally designed to evaluate a single intervention—to complex, multi-drug combinations across distinct patient subpopulations using observational EHR data.

Methods

The authors base their framework on the Rubin-Neyman potential outcomes model, using the Average Treatment response on the Treated (ATT) as their primary metric to isolate true causal impact.

The core methodological innovations include:

  1. Closed Intervention Sets: A novel semantic definition for multi-drug treatments. An intervention set is considered "closed" if dropping any single drug from the mix reduces the overall treatment efficacy. This ensures that the framework only discovers lean, effective drug combinations without unnecessary components.
  2. Advanced Pruning Strategies: To handle the combinatorial explosion of searching through thousands of disease and treatment subsets, the authors introduce Potential Outcome Support Pruning (POSP) and Effective Causal Rule Pruning (ECRP). POSP evaluates whether a causal effect is mathematically estimable by verifying that a subpopulation contains a sufficient sample size of both treated and untreated individuals.
  3. Causal Estimation Benchmarking: The authors implement and compare five distinct statistical methods to estimate counterfactual outcomes: Counterfactual Confidence (CC), Direct Adjustment (DA), Counterfactual Model (CM), Propensity Score Matching (PSM), and Stratified Non-Parametric (SN) estimation.

Datasets

The framework was evaluated using three distinct data sources:

  • Sepsis Cohort: A real-world dataset of 177 adult patients with severe sepsis or septic shock from a Midwest health system, used to evaluate compliance with Surviving Sepsis Campaign (SSC) guidelines.
  • Mayo Clinic Cohort: A large-scale real-world dataset containing 52,139 patients followed over 13 years (1999–2013) to study the 5-year incidence of Type-II Diabetes Mellitus (T2DM) relative to statin (cholesterol drug) use.
  • Synthetic Datasets: Five simulated datasets (5,000 observations each) engineered with specific, known biases (such as direct effects, indirect effects, and complex confounding) to mathematically verify the accuracy of the five estimation methods.

Key Insights and Ideas

  • Pruning via Estimability: Instead of just pruning based on raw frequency thresholds (the standard Apriori approach), POSP prunes based on mathematical estimability. If a subpopulation lacks enough untreated controls to calculate a reliable counterfactual, the algorithm stops exploring that branch entirely.
  • Resolving Clinical Controversies: The framework demonstrates how "contradictory" medical literature can simply be a symptom of unaddressed disease heterogeneity. For instance, the medical community has debated whether statins increase the risk of diabetes. This framework resolves the controversy by isolating specific subpopulations: statins are slightly detrimental to patients with mild hyperlipidemia but are actually protective against diabetes progression for patients with severe hyperlipidemia.

Key Results

  • Pruning Efficiency: The combination of POSP and ECRP dramatically reduced the computational search space compared to traditional support-based pruning, making combinatorial causal mining computationally feasible.
  • Sepsis Guidelines: In the sepsis cohort, translating clinical guidelines into computable rules (shown in Figure 1) and applying the framework revealed that compliance with the "Ventilator rule" (Recommendation 15) significantly reduced inpatient mortality and respiratory complications.
Figure 1: SSC rules for measuring guideline compli- ances
  • Statin Analysis: In the Mayo Clinic dataset, naive counterfactual confidence erroneously showed statins as universally detrimental. However, advanced methods like PSM and Stratified Non-Parametric estimation revealed that for the severe hyperlipidemia subpopulation (HLP2), statins successfully reduced the risk of T2DM progression (yielding negative ATT values ranging from -0.021 to -0.048).
Figure 3: Distribution of the propensity scores be- tween exposed and unexposed groups for the out- come Death when patients and the SSC recommen- dation was Ventilator..

Limitations and Future Work

  • Data Granularity and Timing: The EHR data lacked precise timestamps for certain ICD-9 diagnoses. Consequently, the authors could not always determine if a guideline intervention occurred right before or after a complication arose, potentially underestimating treatment benefits.
  • Workflow Context: The current model evaluates individual rules in isolation rather than modeling the entire clinical workflow sequence. The authors note that future work will focus on capturing the full sequential workflow of clinical guidelines.
  • The Curse of Dimensionality: While the Stratified Non-Parametric method is highly robust against model misspecification, it suffers severely from data fragmentation when stratifying across many variables, leading to empty or tiny patient strata.

Summary updated

Abstract

Our aging population increasingly suffers from multiple chronic diseases simultaneously, necessitating the comprehensive treatment of these conditions. Finding the optimal set of drugs for a combinatorial set of diseases is a combinatorial pattern exploration problem. Association rule mining is a popular tool for such problems, but the requirement of health care for finding causal, rather than associative, patterns renders association rule mining unsuitable. To address this issue, we propose a novel framework based on the Rubin-Neyman causal model for extracting causal rules from observational data, correcting for a number of common biases. Specifically, given a set of interventions and a set of items that define subpopulations (e.g., diseases), we wish to find all subpopulations in which effective intervention combinations exist and in each such subpopulation, we wish to find all intervention combinations such that dropping any intervention from this combination will reduce the efficacy of the treatment. A key aspect of our framework is the concept of closed intervention sets which extend the concept of quantifying the effect of a single intervention to a set of concurrent interventions. We also evaluated our causal rule mining framework on the Electronic Health Records (EHR) data of a large cohort of patients from Mayo Clinic and showed that the patterns we extracted are sufficiently rich to explain the controversial findings in the medical literature regarding the effect of a class of cholesterol drugs on Type-II Diabetes Mellitus (T2DM).
BibTeX
@misc{yadav2016causal,
  author = {Pranjul Yadav and Lisiane Prunelli and Alexander Hoff and Michael Steinbach and Bonnie Westra and Vipin Kumar and Gyorgy Simon},
  title = {{Causal Inference in Observational Data}},
  year = {2016},
  eprint = {1611.04660},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1611.04660},
}
Figure