Practical approach to determine sample size for building logistic prediction models using high-throughput data.

Area of Science:

• Statistics
• Bioinformatics
• Computational Biology

Background:

• Traditional methods for sample size determination in prediction models, like bootstrapping and full permutations, are computationally expensive.
• Overfitting during cross-validation is a significant challenge when building prediction models, especially from microarray data.

Purpose of the Study:

• To propose an efficient empirical method for sample size determination in prediction model development.
• To reduce the high computational cost associated with traditional permutation and bootstrapping methods.

Main Methods:

• Utilized a single representative null distribution instead of computationally intensive full permutations.
• Validated the method using both simulated datasets (with zero effect size) and real data.
• Employed pilot data generated by random sampling from real data for sample size determination.

Main Results:

• The empirical Type I error was confirmed to approach 0.05, indicating reliability.
• The proposed method drastically reduces computation time, with sample size estimation taking approximately 30 minutes.
• Results were comparable to the full permutation method, demonstrating efficacy.

Conclusions:

• The proposed method offers a computationally efficient and reliable approach for sample size determination in prediction models.
• This method can be confidently applied to mitigate overfitting in cross-validation for high-throughput data.
• Facilitates efficient design of clinical studies by reducing time and computational burden.