Inference for change-plane regression

Chaeryon Kang, Hunyong Cho, Rui Song, Moulinath Banerjee, Eric B. Laber, Michael R. Kosorok

Abstract

A key challenge in analyzing the behavior of change-plane estimators is that the objective function often possesses multiple minimizers. This paper proposes two specific estimators—the mean-midpoint and mode-midpoint estimators—to address this non-uniqueness. For each estimator, we establish an $n$-rate of convergence and derive the corresponding limiting distribution. Based on these theoretical foundations, we introduce a parametric bootstrap procedure for statistical inference. The validity of our results and the finite-sample performance of the bootstrap are demonstrated through simulation experiments and an application to latent subgroup identification in the ACTG175 AIDS study.

Introduction

In many scientific fields, particularly precision medicine, researchers often encounter populations that are not homogeneous. Instead, these populations may be split into subgroups where the relationship between a treatment and an outcome shifts abruptly across a hidden boundary. Change-plane regression is a powerful tool for modeling these shifts by partitioning the data space with a hyperplane—a flat boundary in multiple dimensions—and applying distinct linear models to each side.

Despite its utility, change-plane regression has historically been difficult to analyze rigorously. Two primary obstacles exist:

  1. Non-smoothness: The abrupt "jump" at the boundary creates a jagged objective function, rendering standard calculus-based statistical methods (like M-estimation) inapplicable.
  2. Non-uniqueness: There is often no single "best" boundary; rather, a set of multiple boundaries may fit the data equally well.

This paper provides a new theoretical framework to overcome these hurdles, moving change-plane regression from a heuristic optimization problem to a rigorous tool for statistical inference.

Methodology

Addressing Non-Uniqueness

To resolve the problem of multiple optimal solutions, the authors propose two new estimators: the mean-midpoint and mode-midpoint estimators. Rather than selecting an arbitrary boundary from the set of optimal solutions, these estimators systematically identify a representative boundary. The mode-midpoint estimator, in particular, utilizes logic similar to Support Vector Machines (SVMs) by seeking the boundary that maximizes the margin between potential models, leading to a more stable fit.

Convergence and Inference

A significant theoretical contribution of this work is the proof of an $n$-rate of convergence for the boundary parameters. In standard regression, precision typically increases at a $\sqrt{n}$ rate (requiring four times the data to double precision). In change-plane regression, the "sharpness" of the boundary allows the precision to grow linearly with the sample size.

Because the limiting distribution of these estimators involves complex compound Poisson processes, the authors developed a parametric bootstrap procedure. This allows practitioners to generate reliable confidence intervals and p-values for the boundary's location, which was previously a significant challenge in the field.

Computational Optimization

To implement these estimators, the authors utilize Mixed Integer Programming (MIP) and block coordinate descent. These methods allow for efficient searching of the non-convex space to find the optimal hyperplane even in complex datasets.

Experimental Results

Simulation Studies

Extensive Monte Carlo simulations were conducted to verify the theoretical findings. The results confirmed that while regression slopes follow the traditional $\sqrt{n}$-rate, the boundary parameters indeed achieve the faster $n$-rate of convergence. Furthermore, the proposed parametric bootstrap demonstrated excellent finite-sample performance, achieving approximately 95% coverage in simulated scenarios.

Application: ACTG175 AIDS Study

The methodology was applied to data from the ACTG175 clinical trial, which involved 1,046 HIV-infected adults. The goal was to identify if specific patient subgroups responded differently to drug regimens (e.g., ZDV vs. ZDV+ddI) based on age and behavioral factors.

The model identified a distinct subgroup comprising roughly 55% of the participants, defined by age and homosexual activity. In this specific subgroup, the treatment effect of ZDV+ddI was significantly higher, showing an increase of approximately 70 in CD4 cell count compared to the other subgroup. The 95% confidence interval for this effect was [37.7, 103.1], providing clear evidence for targeted treatment strategies.

Discussion and Conclusion

This research transforms change-plane regression into a rigorous framework for subgroup analysis. By establishing the $n$-rate of convergence and providing a valid bootstrap for inference, the authors enable researchers to calculate reliable confidence intervals for "sharp" boundaries in multidimensional data.

Limitations and Future Work

While this work represents a major step forward, some limitations remain:

  • Dimensionality: Global optimization remains computationally intensive for very high-dimensional spaces (typically $p > 20$).
  • Linearity: The current model assumes a flat (hyperplane) boundary; future research may explore "change-manifold" regression for curved boundaries.
  • Multiple Splits: The current framework focuses on a single split into two subgroups. Extending this to multi-threshold models with three or more groups is a promising direction for future study.

Summary updated

Abstract

A key challenge in analyzing the behavior of change-plane estimators is that the objective function has multiple minimizers. Two estimators are proposed to deal with this non-uniqueness. For each estimator, an n-rate of convergence is established, and the limiting distribution is derived. Based on these results, we provide a parametric bootstrap procedure for inference. The validity of our theoretical results and the finite sample performance of the bootstrap are demonstrated through simulation experiments. We illustrate the proposed methods to latent subgroup identification in precision medicine using the ACTG175 AIDS study data.
BibTeX
@inproceedings{kang2025inference,
  author = {Chaeryon Kang and Hunyong Cho and Rui Song and Moulinath Banerjee and Eric B. Laber and Michael R. Kosorok},
  title = {{Inference for change-plane regression}},
  booktitle = {Bernoulli},
  year = {2025},
  doi = {10.3150/24-bej1812},
}
Figure