Locally Weighted Score Estimation for Quantile Classification in Binary Regression Models.

Area of Science:

• Statistics
• Biostatistics
• Machine Learning

Background:

• Binary response regression is widely used for classification tasks.
• Existing methods often ignore application-specific probability thresholds.
• Maximum likelihood estimation is a common but potentially suboptimal approach.

Purpose of the Study:

• To develop a novel estimation procedure for linear logistic models that incorporates a priori probability thresholds.
• To improve classification accuracy, particularly for high- and low-risk groups.
• To reduce prediction error rates in binary response regression.

Main Methods:

• A locally weighted score equation approach using a kernel-like weight function centered at the threshold.
• Cross-validation of a hybrid loss function combining classification error and divergence.
• Exploration of alternative cross-validation functions based on common binary classification metrics.

Main Results:

• The proposed method demonstrates reduced error rates compared to maximum likelihood estimation.
• Effectiveness is particularly notable under certain forms of model misspecification.
• Simulations and a melanoma dataset analysis validate the practical utility of the method.

Conclusions:

• Incorporating application-specific thresholds into the estimation procedure enhances classification performance.
• The locally weighted approach offers a robust alternative for predictive modeling in binary response regression.
• This method provides a valuable tool for risk stratification and classification in various applications.