Abstract
From personalized medicine to targeted advertising, providing a sequence of decisions based on historical covariates and outcome data is a critical task. This requires a deep understanding of both the dynamics and heterogeneity of treatment effects. This paper addresses the challenge of detecting abrupt changes in the Conditional Average Treatment Effect (CATE) in a sequential, online fashion. We propose a nonparametric model to allow for maximal flexibility and robustness, accounting for temporal dependence on historical covariates and noise. We introduce a kernel-based change point estimator and demonstrate its consistency in terms of detection delay under average run length control. Numerical results support our theoretical findings.
Introduction
In real-world decision-making—such as adjusting clinical treatments or optimizing targeted advertising—decisions are made sequentially based on an individual's history. Two primary challenges complicate this process: heterogeneity and dynamics. Heterogeneity refers to the fact that treatment effects vary across individuals based on covariates (e.g., age, purchase history). Dynamics refers to the fact that these relationships are not static; a patient's response to a drug or a consumer's reaction to an ad can shift abruptly over time.
Existing statistical tools often struggle to handle both heterogeneity and temporal dynamics simultaneously, particularly when data points are not independent over time. This research fills that gap by providing a framework to detect exactly when treatment effects change abruptly, enabling rapid strategic adjustments in real-time monitoring settings.
Methodology: Online CATE Change Point Detection
The core of this work is the Online CATE Change Point Detector (Algorithm 1), which operates within the potential outcome framework (Neyman-Rubin model).
Nonparametric Modeling and Direct Estimation
Rather than assuming a rigid linear relationship, the authors employ Nadaraya–Watson kernel estimators. A key innovation is the "One-K" approach. Traditional methods ("Two-K") estimate the mean functions for treated and control groups separately and then subtract them. In contrast, the "One-K" method estimates the CATE function directly. This approach requires fewer tuning parameters and proves significantly more robust when individual mean functions are "rough" or oscillatory, but their difference (the treatment effect) remains smooth.
Sequential Monitoring and Temporal Dependence
The algorithm monitors incoming data using a sliding window strategy, comparing two adjacent time windows. If the estimated treatment effects between these windows differ beyond a calculated threshold ($\varepsilon$), a change point is flagged. Crucially, the framework accounts for $\alpha$-mixing, allowing it to model short-range temporal dependence rather than assuming independent and identically distributed (i.i.d.) data, which is more realistic for time-series applications.
Experimental Evaluation
The proposed method was evaluated across four generative simulation scenarios designed to test various complexities:
- Scenario 1: A randomized experiment with simple mean and treatment functions.
- Scenario 2: Non-constant propensity scores (probability of treatment) and temporal noise.
- Scenario 3: Highly oscillatory mean functions to test the limits of nonparametric estimation.
- Scenario 4: High-dimensional settings ($d=6$) with complex covariate interactions and temporal dependence.
Key Results
- Theoretical Consistency: The authors prove that the estimator is consistent, meaning detection delay becomes negligible relative to total time as the pre-change sample size grows.
- Performance: In simulations, the proposed Algorithm 1 consistently outperformed difference-based kernel (DK) competitors, particularly in Scenarios 3 and 4 where individual mean functions were difficult to estimate.
- Robustness: The algorithm maintained performance across different window widths and covariate dimensions, demonstrating its versatility.
Discussion and Future Work
The "One-K" estimation strategy represents a significant advancement in CATE estimation, particularly for "messy" data where individual group responses are noisy but the treatment effect itself is stable.
However, some limitations remain. The current model assumes a constant propensity score over time for simplicity. Future work could extend this to fully dynamic propensity score models. Additionally, while the model handles moderate dimensionality ($d=6$), truly high-dimensional settings would benefit from integrating regularization techniques like Lasso or Ridge regression. Finally, exploring more general function classes for CATE, such as those with bounded variation, could further enhance the model's flexibility.
Summary updated