Introduction
In causal inference, the gold standard for measuring intervention effectiveness is a single dataset containing covariates, treatment assignment, compliance, and outcomes. However, in many real-world scenarios—such as "Online-to-Offline" (O2O) marketing—data is fragmented. A researcher might know who received an online coupon (Dataset A) and who made a purchase in a physical store (Dataset B), but privacy or technical silos prevent linking these individuals. Furthermore, noncompliance (where individuals do not follow their assigned treatment) necessitates estimating the Local Average Treatment Effect (LATE) rather than a simple average effect.
The Two-Regime Design (2RD)
To address these challenges, the authors propose a Two-Regime Design (2RD) framework. This framework allows for the identification of LATE even when treatment status and outcomes are never observed together in the same dataset. The process involves a regime assignment (e.g., experimental vs. natural state) followed by treatment assignment within those regimes.
Directly Weighted Least Squares (DWLS)
The core technical innovation is the Directly Weighted Least Squares (DWLS) estimator. Traditional methods like Inverse Probability Weighting (IPW) often suffer from high variance when propensity scores are near zero. DWLS overcomes this by using the Propensity-Score-Difference (PSD) directly as a weight.
Unlike previous "direct" methods that relied on unstable minimax optimization, DWLS transforms the problem into a standard weighted least squares task. This allows researchers to use familiar model selection techniques like cross-validation, significantly improving practical usability and numerical stability.
Experimental Results
The authors evaluated DWLS against several baselines, including Inverse Weighted Least Squares (IWLS) and minimax-based Direct Least Squares (DLS).
Synthetic Benchmarks
In synthetic tests with varying treatment effect shapes (constant, linear, and logistic) and covariate dimensions, DWLS demonstrated superior stability. As shown in the performance table below, DWLS consistently achieved lower error rates, particularly as the complexity of the data (number of covariates $q_x$) increased. While other methods like IWLS often "exploded" with high variance when propensity scores were near zero, DWLS remained robust.
Real-World Validation: The JTPA Study
The method was also applied to the National Job Training Partnership Act (JTPA) dataset, a large-scale randomized trial. DWLS achieved a Root Mean Squared Error (RMSE) of $252 \pm 141$, proving far more stable than the DLS method ($946 \pm 663$) and competitive with more complex weighting schemes.
Conclusion
DWLS offers a robust and practical solution for estimating causal effects in fragmented data environments. By relaxing the "positivity" assumption and providing a stable, easily tunable optimization objective, it bridges the gap between theoretical causal inference and the messy reality of siloed data. Future work will focus on extending this framework to handle time-varying covariates and heterogeneous populations across datasets.
Summary updated