Unimodal regression using Bernstein-Schoenberg splines and penalties

Area of Science:

• Statistics
• Nonparametric Regression
• Shape Constrained Regression

Background:

• While nonparametric shape constrained regression is well-researched, methods for unimodality are scarce.
• Unimodal regression is crucial for applications such as dose-response analysis.

Purpose of the Study:

• To propose novel unimodal spline regression methods.
• To enhance the estimation of functional relationships using unimodality constraints.

Main Methods:

• Utilizing Bernstein-Schoenberg splines for their shape preservation properties.
• Employing penalized splines to achieve unimodal and smooth solutions.
• Extending penalized splines to penalize against general parametric functions.
• Developing restricted maximum likelihood and Bayesian approaches for parameter selection under unimodality constraints.

Main Results:

• The proposed unimodal spline regression methods were compared to existing approaches via simulation.
• Application to a dose-response dataset demonstrated the practical utility of the methods.
• Results indicate substantial improvements in functional relationship estimation when unimodality constraints are applied.

Conclusions:

• The proposed unimodal spline regression methods offer significant advantages over traditional approaches.
• The combination of unimodality constraints and penalties effectively improves estimation accuracy.
• These methods provide valuable tools for applications requiring unimodal function estimation.