Sample size determination for prediction models via learning-type curves.

Area of Science:

• Statistics
• Machine Learning
• Biostatistics

Background:

• Sample size determination is crucial for the reliability of prediction models.
• Existing methods often lack robustness and efficiency, especially when dealing with limited data or extrapolating findings.

Purpose of the Study:

• To develop and evaluate novel methodologies for sample size determination in prediction modeling.
• To enhance the performance and statistical efficiency of sample size calculations by leveraging learning curves.

Main Methods:

• Proposing two methods: a deterministic learning curve skeleton and a Gaussian process model built upon it.
• Utilizing various learning algorithms for primary endpoint modeling and distinct efficacy measures.
• Illustrating the methods with binary and survival endpoints.

Main Results:

• Combining individual sample size calculations via learning curves universally improves performance.
• The Gaussian process-based learning curve demonstrates superior robustness and statistical efficiency.
• Computational efficiency is comparable between the proposed methods.

Conclusions:

• Learning curves effectively integrate information across different sample sizes for more reliable sample size determination.
• Anchoring sample size extrapolations against historical data is recommended when available.
• The Gaussian process approach offers a statistically sound and efficient solution for sample size planning in prediction modeling.