Building a better model: abandon kitchen sink regression.

Area of Science:

• Biostatistics
• Epidemiology
• Medical Research Methodology

Background:

• Multivariable regression models are crucial in medical research for understanding disease risk factors.
• Variable selection is a critical step in regression modeling, influencing the validity of results.
• Current practices like 'kitchen sink regression' often rely on statistical criteria rather than causal reasoning.

Purpose of the Study:

• To critically evaluate the 'kitchen sink regression' method for variable selection.
• To identify the inherent pitfalls and limitations of using p-values or information criteria for model building.
• To propose alternative, more robust methods for regression model development in medical research.

Main Methods:

• Critical examination of the 'kitchen sink regression' approach.
• Illustration of pitfalls using examples from perinatal/neonatal medicine.
• Introduction and application of Directed Acyclic Graphs (DAGs) for causal analysis.

Main Results:

• 'Kitchen sink regression' disregards variable directionality, lacks causal interpretation, inflates Type I error rates, risks overfitting, and ignores content expertise.
• Five key issues identified with the 'kitchen sink regression' method.
• DAGs provide a framework for understanding and visualizing causal relationships.

Conclusions:

• 'Kitchen sink regression' is a flawed method for variable selection in medical research.
• Directed acyclic graphs (DAGs) are recommended to guide variable selection for examining risk factor-outcome associations.
• A more thoughtful and informed approach to regression modeling, incorporating causal reasoning, is essential.