Date of Award

Spring 4-22-2026

Embargo Period

4-30-2028

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Public Health Sciences

College

College of Graduate Studies

First Advisor

Bethany Wolf

Second Advisor

Paul Nietert

Third Advisor

Diane Kamen

Fourth Advisor

Jim Oates

Fifth Advisor

Jaime Speiser

Sixth Advisor

Jonathan Beall

Abstract

Dichotomization of continuous predictors to predict binary outcomes is a prevalent practice in clinical settings, utilized for patient diagnosis, prognosis, and resource allocation. In order to make decisions that are binary in nature, clinicians need binary interpretations on continuous measures. Despite its widespread use in clinical studies, dichotomizing continuous predictors to discriminate binary outcomes has faced substantial criticism since it can lead to a loss of valuable information and statistical power. Given the prevalent use of dichotomization in clinical decision making, outright dismissal based solely on the loss of information and power might not be the most prudent approach. There are scenarios where the information loss using a dichotomized version of a continuous predictor in regression estimation is not greater than using the original continuous version of the predictor.

By extending the model to incorporate multiple predictors, developing clinical prediction models often involves balancing uncertainty in the prediction of patient outcomes with the desire for interpretable, actionable models. Incorporating prior clinical knowledge and existing evidence when constructing these models can improve prediction accuracy and enhance clinical relevance. Traditional decision tree models provide an intuitive and interpretable approach for clinical research; however, they do not have an inherent mechanism to capture uncertainty in predictions. While methods based on sampling variability, such as bootstrapping or binomial confidence intervals, can be used with traditional decision trees to understand the uncertainty around predictions, these methods only approximate sampling variability rather than integrating uncertainty across parameters and possible model configurations. To address the limitations in traditional decision tree approaches, we develop a Bayesian decision tree approach to quantify parameter and structural uncertainty through a coherent probabilistic framework and to allow for integration of prior knowledge. The proposed approaches, known as BayLeaf-Reg and BayLeaf-Bin, provides probabilistic inference on the tree structure by estimating posterior distributions of variables being split on and their cut-points, yields posterior summaries of predictor importance and predicted outcome probabilities, and does not suffer from the same weak learner tendencies as traditional decision trees.

Rights

Copyright is held by the author. All rights reserved.

Available for download on Sunday, April 30, 2028

Included in

Biostatistics Commons

Share

COinS