Decision trees in epidemiological research.

Area of Science:

• Biostatistics
• Data Science
• Population Health

Background:

• Identifying homogeneous population subgroups is crucial for understanding effect mechanisms and designing targeted interventions.
• Standard statistical methods often present challenges in effectively identifying these relevant subgroups.

Purpose of the Study:

• To review and compare decision tree techniques for population subgroup identification.
• To assess the performance of Classification and Regression Trees (CART) and Conditional Inference Trees (CTree).
• To introduce a novel visualization method for decision tree-derived subgroups.

Main Methods:

• Literature review of decision tree methods for population partitioning based on covariates.
• Comparative analysis of CART and CTree using simulation studies and real-world data (Box Lunch Study).
• Development of a novel graphical visualization technique for decision tree subgroups.

Main Results:

• Both CART and CTree successfully identify homogeneous subgroups and improve prediction accuracy compared to traditional regression.
• CTree employs a formal statistical hypothesis testing framework, simplifying tree building and interpretation.
• The novel visualization method offers a more scientifically meaningful characterization of identified subgroups.

Conclusions:

• Decision trees are valuable for identifying subgroups based on combined individual characteristics.
• CTree is advocated for its superior simplicity and interpretability in subgroup analysis.
• Enhanced subgroup visualization aids in scientific interpretation and application.