Why significant variables aren't automatically good predictors.

Area of Science:

• Genetics
• Biostatistics
• Computational Biology

Background:

• Genome-wide association studies (GWAS) identify statistically significant variants but often fail to yield strong predictive models for complex diseases.
• The disconnect between statistical significance and predictive accuracy is a persistent challenge in various scientific fields.

Purpose of the Study:

• To investigate why statistically significant variants from GWAS do not translate to effective prediction for personalized medicine.
• To propose and evaluate alternative methods for identifying variables with high predictive value.

Main Methods:

• Statistical analysis of significance versus predictivity.
• Simulations to illustrate the relationship between significance and prediction.
• Application of the partition retention method to a breast cancer dataset.

Main Results:

• Highly significant variables do not necessarily equate to high predictive power.
• Predictive variables may not always achieve high statistical significance, thus being overlooked by standard methods.
• The partition retention method significantly reduced classification error in a breast cancer dataset from 30% to 8%.

Conclusions:

• Rethinking variable selection criteria beyond statistical significance is crucial for advancing predictive modeling in complex diseases.
• Novel methods like the partition retention method show promise for identifying truly predictive variables.
• Improved prediction accuracy holds potential for enhancing personalized medicine strategies.