Modelling survival data to account for model uncertainty: a single model or model averaging?

Area of Science:

• Biostatistics
• Genomics
• Cancer Research

Background:

• Predicting patient survival is crucial in cancer research.
• Traditional methods often select a single
• best
• statistical model based on data fit.
• Model uncertainty can impact the reliability of survival predictions.

Purpose of the Study:

• To evaluate Bayesian model averaging (BMA) for predicting lymphoma cancer survival.
• To compare BMA with single model selection approaches.
• To investigate the influence of sample size on model selection and prediction accuracy.

Main Methods:

• Utilized three survival models: single Weibull, mixture of Weibulls, and a cure model.
• Employed Bayesian model averaging (BMA) to handle model uncertainty.
• Applied the Bayesian information criterion (BIC) for model comparison.

Main Results:

• With large sample sizes, a single model often emerges as superior based on BIC.
• Reduced sample sizes obscured the superiority of any single model.
• BMA provided robust predictions, especially when sample size was limited.

Conclusions:

• BMA is a valuable approach for cancer survival analysis, particularly when model uncertainty is high or sample sizes are small.
• BMA can yield more reliable predictions than single-model approaches in certain scenarios.
• This method facilitates deeper understanding of gene expression's role in patient survival.