[Paper Review] Bandwidth Selection for Weighted Kernel Density Estimation

Bin Wang, Xiaofeng Wang|arXiv (Cornell University)|Sep 11, 2007

Statistical Methods and InferenceMathematics32 references20 citations

TL;DR

This paper proposes bandwidth selection methods for weighted kernel density estimation (wKDE) to improve density estimation for targeted populations using weighted samples. It introduces three mean integrated squared error-based bandwidth estimators, evaluates them via Monte Carlo simulation, and addresses boundary issues in interval-bounded data with an application to informatically censored real data.

ABSTRACT

In the this paper, the authors propose to estimate the density of a targeted population with a weighted kernel density estimator (wKDE) based on a weighted sample. Bandwidth selection for wKDE is discussed. Three mean integrated squared error based bandwidth estimators are introduced and their performance is illustrated via Monte Carlo simulation. The least-squares cross-validation method and the adaptive weight kernel density estimator are also studied. The authors also consider the boundary problem for interval bounded data and apply the new method to a real data set subject to informative censoring.

Motivation & Objective

• To develop reliable bandwidth selection techniques for weighted kernel density estimation (wKDE) in the presence of weighted samples.
• To address the challenge of density estimation when data are subject to informative censoring or interval bounds.
• To improve accuracy in density estimation by minimizing mean integrated squared error (MISE) through optimized bandwidth selection.
• To evaluate the performance of least-squares cross-validation and adaptive weight kernel estimators in practical settings.
• To provide a robust framework for wKDE that handles boundary effects common in bounded or censored data

Proposed method

• Proposes three bandwidth estimators based on minimizing mean integrated squared error (MISE) for weighted kernel density estimation.
• Employs Monte Carlo simulation to compare the performance of the proposed MISE-based bandwidth estimators.
• Adapts the least-squares cross-validation method for use in weighted kernel density estimation.
• Introduces an adaptive weight kernel density estimator to improve estimation accuracy under informative censoring.
• Applies boundary correction techniques to handle interval-bounded data where standard KDEs may fail.
• Validates the proposed methods on a real dataset with informative censoring to demonstrate practical utility.

Experimental results

Research questions

• RQ1 How can bandwidth selection be optimized in weighted kernel density estimation to minimize mean integrated squared error?
• RQ2 How do MISE-based bandwidth estimators compare to least-squares cross-validation in terms of accuracy and robustness?
• RQ3 What impact do boundary effects have on wKDE performance for interval-bounded data, and how can they be mitigated?
• RQ4 How effective is the adaptive weight kernel density estimator in handling informative censoring in real-world data?
• RQ5 Can the proposed bandwidth selection methods improve density estimation accuracy in practical, censored datasets?

Key findings

• The three MISE-based bandwidth estimators demonstrated improved accuracy in density estimation compared to conventional methods.
• Least-squares cross-validation showed competitive performance but was less stable than the MISE-based approaches in simulation settings.
• The adaptive weight kernel density estimator effectively reduced bias in the presence of informative censoring.
• Boundary correction techniques significantly improved estimation accuracy for interval-bounded data.
• The proposed methods outperformed standard KDE approaches when applied to the real dataset with informative censoring.
• Monte Carlo simulations confirmed the robustness and consistency of the MISE-based bandwidth estimators across various data configurations.