Change-point models for hazard functions

Hans‐Georg Müller, J.-l. WangLecture Notes-monograph Series 1994 · DOI · OpenAlex

Abstract

A review is presented of parametric and nonparametric models and corresponding estimation procedures for change-points in hazard functions where the data are possibly subject to random censoring. In particular, we discuss nonparametric models and the application of nonparametric smoothing techniques for change-point estimation and estimation of a hazard function when a change-point is present. Preliminary theoretical results are mentioned and a simulation study provides further insight.

Introduction

In survival analysis, the hazard function represents the instantaneous risk of an event—such as a medical complication, machine failure, or mortality—occurring at a specific time $t$. In many practical applications, this risk is not static. For instance, in bone marrow transplant patients, the risk of complications is often extremely high immediately following the procedure before stabilizing into a lower, long-term risk profile.

The moment this risk profile shifts is known as the change-point ($\tau$). Identifying this point is critical for clinical decision-making and reliability engineering. However, detecting these shifts is complicated by random censoring, where a study ends or a subject drops out before the event of interest occurs. This paper reviews the evolution of models designed to pinpoint these shifts, moving from rigid parametric assumptions to flexible nonparametric kernel methods.

Modeling Frameworks

The literature generally categorizes change-point problems into three distinct modeling approaches, each offering different trade-offs between simplicity and realism.

1. Parametric Models (Model 1)

This is the classical approach, often referred to as a "step-function" model. It assumes the hazard rate is a constant $\lambda_1$ before the change-point and shifts to a different constant $\lambda_2$ afterward. While mathematically tractable and easy to interpret, this model is often too brittle for biological data where the underlying risk may fluctuate even outside the primary change-point.

2. Nonparametric Discontinuity (Model 2)

This model allows for a more flexible hazard function that is generally smooth but contains a sharp "jump" or discontinuity at the change-point $\tau$. This is useful for modeling events where a specific intervention or biological threshold causes an immediate and sustained shift in risk.

3. Smooth Approximation (Model 3)

In some scenarios, a literal "jump" is biologically implausible. Instead, the change-point is defined as the moment of most rapid change—mathematically represented as the peak of the hazard function's first derivative. This approach treats the change-point as a transition zone rather than a discrete break.

Methodology: Kernel Smoothing and One-Sided Lenses

The core technical challenge in change-point estimation is avoiding the "smoothing trap." Standard smoothing techniques (like simple moving averages) tend to blur discontinuities, averaging out the very signal researchers are trying to detect.

To solve this, the authors highlight the use of Kernel Smoothing with specialized "one-sided kernels." This method evaluates the data through two different mathematical "lenses":

  • One lens looks exclusively at data to the left of a potential time point.
  • One lens looks exclusively to the right.

When the difference between these two views is maximized, the algorithm identifies a likely change-point. Furthermore, boundary kernels are employed to ensure that the estimation remains accurate near the edges of the data range, preventing the "smearing" of the hazard rate near the change-point.

Empirical Evidence and Simulation

The effectiveness of these models was tested using both historical medical data and Monte Carlo simulations.

Medical Datasets

  • Leukemia Study: Analysis of 84 acute lymphoblastic leukemia patients (Matthews and Farewell, 1982) to identify shifts in remission risk.
  • Lymphoma Study: Data on 31 advanced non-Hodgkin’s lymphoma patients. Both datasets involve significant censoring, requiring the use of modified log-likelihood functions to account for incomplete timelines.

Simulation Results

Using 200 Monte Carlo runs with $n=200$ observations, the study revealed several key insights:

  • Model Adaptivity: Kernel methods are remarkably robust. They can adapt to both sharp jumps (Model 2) and rapid smooth shifts (Model 3) without prior knowledge of the underlying structure.
  • Kernel Selection: A specific polynomial construction labeled $K_{-3}$ was found to be the most robust. This kernel places higher weight on data points closest to the potential change-point, improving the precision of the estimate.
  • Bandwidth Sensitivity: The "bandwidth" (the size of the data window the kernel considers) is the most critical parameter. A bandwidth that is too narrow introduces noise, while one that is too wide obscures the change-point.

Conclusion and Future Directions

While nonparametric kernel methods offer a powerful toolkit for risk analysis, several challenges remain. The selection of an optimal bandwidth currently lacks a universal, automated "plug-in" method, often requiring manual tuning. Additionally, while the preliminary asymptotic theory for these models is promising, further research is needed to fully characterize their behavior under complex random censoring patterns. Despite these hurdles, the shift toward flexible, kernel-based models represents a significant step forward in accurately identifying the "Aha!" moments in risk trajectory.

Summary updated

Abstract

A review is presented of parametric and nonparametric models and corresponding estimation procedures for change-points in hazard functions where the data are possibly subject to random censoring. In particular, we discuss nonparametric models and the application of nonparametric smoothing techniques for change-point estimation and estimation of a hazard function when a changepoint is present. Preliminary theoretical results are mentioned and a simulation study provides further insight.
BibTeX
@inproceedings{muller1994change,
  author = {Hans‐Georg Müller and J.-L. Wang},
  title = {{Change-point models for hazard functions}},
  booktitle = {Lecture Notes-monograph Series},
  year = {1994},
  doi = {10.1214/lnms/1215463127},
}
Figure