Estimating Clustered Data Causal Effects with DiConfounder: A Double‑Difference Framework

The article introduces the concept of clustered data, where samples are grouped (e.g., by city, product category, or hospital) and experiments assign entire clusters to treatment or control, highlighting the need for cluster‑level analysis due to natural grouping, regulatory, ethical, or cost constraints.

Two motivating examples are given: (1) price‑discrimination‑free product pricing experiments, and (2) policy analysis of medical insurance effects where the policy can only be applied at the hospital level, which is the background of the ACIC2022 data challenge.

The ACIC2022 dataset contains four years of medical‑policy data for 3,400 hospitals, each with over one million rows (total >100 GB). The data are highly imbalanced, contain missing values, and exhibit large distributional differences between control and treatment groups, measured by the ASMD metric.

The authors adopt the Rubin Causal Model (RCM) with three core assumptions—no unmeasured confounding, positivity, and SUTVA—modified to a “differential no‑confounding” condition for pre‑ and post‑policy periods. They define three related estimands (yearly SATT, group‑level SATT, hospital‑level SATT) and derive a double‑difference estimator.

To estimate the causal effect, they employ a causal‑forest based R‑Learner (Susan Athey’s method) and design the DiConfounder algorithm, which consists of six steps: (1) feature engineering with hospital‑level and patient‑level attributes, including trends; (2) handling missing data and imputing gaps; (3‑5) training outcome, propensity‑score, and auxiliary models and feeding them into the causal forest; (6) aggregating SATT estimates and quantifying uncertainty via bootstrap or analytical formulas.

The final model achieved a coverage of 80 % (target 90 %) and an RMSE around 10, outperforming the average competition error. Analysis shows the method works best when group differences are small and individual differences are large, and less well when group heterogeneity dominates.

In summary, DiConfounder transforms the estimand into a differential form, applies a double‑robust causal‑forest estimator, and provides uncertainty quantification, offering a practical solution for large‑scale clustered causal inference problems such as healthcare policy evaluation.