Type I error control for tree classification.

Area of Science:

• Statistics
• Machine Learning
• Data Mining

Background:

• Binary tree classification is widely used for population segmentation based on outcome variables and predictors.
• The method is prone to Type I errors due to numerous candidate predictors and cutoff values.
• Existing literature inadequately addresses the multiplicity issue in binary tree classification.

Purpose of the Study:

• To propose a novel binary tree classification method.
• To control the probability of Type I errors in predictor selection.
• To establish a maximum acceptable probability (e.g., 5%) for accepting a predictor.

Main Methods:

• Developing a modified binary tree classification algorithm.
• Implementing a statistical procedure to adjust for multiple testing.
• Simulating performance under various data conditions.

Main Results:

• The proposed method effectively controls the Type I error rate.
• Performance evaluation demonstrates reduced false positive rates compared to standard methods.
• The method maintains reasonable classification accuracy.

Conclusions:

• The new binary tree classification technique offers a robust solution for managing Type I errors.
• This approach enhances the reliability of predictor selection in complex datasets.
• Further research can explore extensions to other classification algorithms.