Order-constrained linear optimization

Area of Science:

• Social Sciences
• Behavioral Sciences
• Health Sciences

Background:

• Ordinal data is prevalent in social, behavioral, and health sciences.
• General linear models (GLM) are common but do not optimally model ordinal properties.
• Existing methods often fail to fully leverage the ordinal nature of data.

Purpose of the Study:

• Introduce an order-constrained linear least-squares (OCLO) optimization algorithm.
• Maximize linear least-squares fit while prioritizing ordinal properties using Kendall's τ.
• Evaluate OCLO's performance against ordinary least squares (OLS) for ordinal data.

Main Methods:

• Developed an order-constrained linear least-squares (OCLO) optimization algorithm.
• Algorithm builds upon maximum rank correlation estimator and general monotone model.
• Analyzed simulated data under various conditions, including fat-tailed distributions.

Main Results:

• OCLO shows minimal bias and variance with minimal loss in predictive accuracy for OLS-adherent data.
• OCLO demonstrates reduced bias and variance with substantially improved predictive accuracy for fat-tailed data.
• OCLO's predictive advantages persist even after outlier removal in skewed datasets.

Conclusions:

• OCLO offers a superior method for analyzing ordinal data in social, behavioral, and health sciences.
• The algorithm effectively handles data with extreme scores, improving predictive modeling.
• OCLO provides a robust alternative to OLS when dealing with the nuances of ordinal data.