Parsimonious covariate selection for a multicategory ordered response.

Area of Science:

• Statistics
• Biostatistics
• Machine Learning

Background:

• Ordinal categorical responses are common in scientific research.
• Analyzing ultrahigh dimensional data presents significant statistical challenges.
• Existing models may not adequately capture unique covariate effects at each ordinal level.

Purpose of the Study:

• To propose a flexible continuation ratio (CR) model for ordinal responses with ultrahigh dimensional data.
• To characterize unique covariate effects at each response level.
• To provide robust methods for model evaluation and interpretation.

Main Methods:

• Developed a flexible CR model based on the logit of conditional discrete hazard functions.
• Proposed two modeling strategies: fixed covariates with varying coefficients, and varying covariates and coefficients.
• Utilized nonparametric bootstrap for prediction error estimation and robust standard error calculation.
• Employed graphical and numerical methods (cumulative sum of residuals) for goodness-of-fit assessment.

Main Results:

• The proposed CR model effectively characterizes unique covariate effects across response levels.
• Bootstrap methods provide reliable estimates of prediction error and standard errors.
• The model allows for flexible covariate selection and coefficient estimation.
• Simulation studies demonstrate good performance in finite samples.

Conclusions:

• The flexible CR model offers a powerful approach for analyzing ultrahigh dimensional ordinal data.
• The methods facilitate improved prediction and interpretation of covariate effects.
• The approach is applicable to real-world datasets, as shown in the B-cell acute lymphocytic leukemia example.