Modeling time-to-event (survival) data using classification tree analysis.

Area of Science:

• Health research methodology
• Biostatistics
• Survival analysis

Background:

• Survival analysis is crucial in health research for studying time-to-event data, including censored observations.
• Cox regression is standard but has assumptions that are often violated.
• There is a need for flexible and robust methods for analyzing censored data.

Purpose of the Study:

• Introduce Classification Tree Analysis (CTA) as a flexible alternative for modeling censored data.
• Compare the performance of CTA survival models against Cox regression and a naive model.
• Evaluate CTA's ability to provide parsimonious, transparent, and accurate predictive models.

Main Methods:

• Identified statistically valid, reproducible, and generalizable CTA survival models using empirical data.
• Compared predictive accuracy of CTA models with Cox regression and an unadjusted naive model.
• Assessed model performance using integrated Brier scores and survival curve comparisons.

Main Results:

• Cox regression models best predict the average incidence of an outcome over time.
• CTA survival models demonstrated superior performance in predicting either relatively high or low incidence of the outcome over time.
• CTA models offer explicit maximization of predictive accuracy and transparency.

Conclusions:

• CTA survival models present significant advantages over Cox regression, including enhanced predictive accuracy, parsimony, and transparency.
• CTA provides statistically robust methods and clear decision rules for analyzing censored data.
• Researchers seeking accurate prognoses and interpretable decision-making tools should consider the CTA-survival framework.