Bootstrap confidence intervals for the optimal cutoff point to bisect estimated probabilities from logistic...

Area of Science:

• Statistics
• Biostatistics
• Machine Learning

Background:

• Optimal cutoff point selection is critical for binary classification using logistic regression models.
• Existing methods for threshold estimation often lack robust statistical inference.
• Accurate classification is vital in fields like disease diagnosis and risk prediction.

Purpose of the Study:

• To develop and evaluate novel bootstrap-based statistical methods for estimating the optimal cutoff point in logistic regression.
• To provide reliable confidence intervals for the estimated optimal threshold.
• To enhance the precision and inferential capabilities in binary classification tasks.

Main Methods:

• Nonparametric bootstrap standard errors and quantile intervals for threshold estimation.
• A refined bagging estimator for the optimal cutoff point.
• Application of the infinitesimal jackknife method for confidence interval construction.

Main Results:

• The proposed bootstrap methods, particularly the bagging estimator, demonstrate strong empirical performance in simulations.
• The infinitesimal jackknife method successfully provides confidence intervals for the optimal cutoff point.
• The methods are validated through practical application on a fertility dataset for seminal quality prediction.

Conclusions:

• The developed bootstrap-based methods offer a statistically sound approach to determining optimal cutoff points in logistic regression.
• These methods improve the reliability and interpretability of classification models by providing confidence intervals.
• The study contributes valuable tools for accurate risk assessment and prediction in various scientific domains.