Sample size requirements are not being considered in studies developing prediction models for binary outcomes: a...

Area of Science:

• Clinical prediction modeling
• Biostatistics
• Health research methodology

Background:

• Appropriate sample size is crucial for developing reliable clinical prediction models.
• This study reviews sample size considerations in models predicting binary outcomes.

Approach:

• Searched PubMed for studies (July 2020) developing prediction models for binary outcomes.
• Reviewed sample size calculations, comparing used vs. required sizes for risk estimation and overfitting minimization.
• Calculated minimum sample sizes and assessed adherence to the 10 events per variable (EPV) rule.

Key Points:

• Only 8% of 119 studies justified sample size.
• 73% used insufficient sample sizes to estimate overall risk and minimize overfitting.
• 75% did not meet the 10 EPV criteria, impacting model performance (median c-statistic 0.80).

Conclusions:

• Clinical prediction models are frequently developed without adequate sample size calculations, compromising risk estimation accuracy.
• Researchers are urged to justify, perform, and report sample size calculations for robust prediction model development.