Abstract:
|
Motivated by an ill-conditioned data set measuring the effects of air pollution on human mortality, the study explores ridge regression and various ridge parameter selection mechanisms, including cross validation and Bayesian approaches, in a general weighted framework. The advantage is that it permits consideration of a range of loss functions, and through a pre-analysis aimed at diagnosing risk even multiple loss functions at the same time. The study provides a framework that synchronizes the most common approaches to ridge parameter selection and enough flexibility to yield a parallel methodology for exponential family regression models.
The development stems from ridge regression's asymptotic connections with classical James-Stein estimation, also treated in a weighted form, and asymptotic evaluation of risk. The extension to non-normal regression uses entropy loss to evaluate estimators and relies heavily on Bayesian hierarchical modeling for their construction.
|