L1 penalized continuation ratio models for ordinal response prediction using high-dimensional datasets.

Area of Science:

• Genomics
• Biostatistics
• Statistical modeling

Background:

• Ordinal response data is common in health outcomes.
• Current analysis of genomic ordinal data uses dichotomous methods, losing power and increasing errors.
• High-throughput genomic datasets require robust analytical approaches.

Purpose of the Study:

• To introduce an innovative frequentist approach for modeling ordinal responses using gene expression microarray data.
• To combine L(1) penalization and continuation ratio models for enhanced analysis.
• To evaluate the performance of computational approaches and model selection criteria.

Main Methods:

• Developed a frequentist L(1) penalized continuation ratio model.
• Conducted simulation studies to assess model performance.
• Applied the model to three microarray gene expression datasets for ordinal classification.

Main Results:

• The L(1) penalized constrained continuation ratio model effectively models ordinal responses.
• This approach is particularly useful when the number of covariates exceeds the sample size (p > n).
• Model selection (AIC vs. BIC) depends on the similarity of underlying disease pathologies.

Conclusions:

• The L(1) penalized constrained continuation ratio model offers a powerful alternative for analyzing genomic ordinal data.
• It overcomes limitations of dichotomous analysis, improving statistical power and accuracy.
• The choice between AIC and BIC for model selection should be guided by biological context.