B1761
Title: Reweighting data in penalized optimization models: An approach to maximize subgroup fairness
Authors: Courtney Paulson - University of New Hampshire (United States) [presenting]
Daniel Smolyak - University of Maryland (United States)
Margret Vilborg Bjarnadottir - University of Maryland (United States)
Abstract: Regularized regression methods have become a popular, nearly ubiquitous tool for approaching high-dimensional data problems. In many fields, however, regularization applies a one-size-fits-all approach to data with subgroups that would benefit from a more tailored estimation. For example, as demonstrated during the COVID-19 pandemic, medical professionals must often develop separate recommendations for high-risk or other diverging groups by identifying factors that lead to different outcomes from thousands of possible patient covariates, including demographics, drug interactions, etc. Regularized methods are ideal for this high-dimensional setup, but traditional regularization estimates only one relationship per covariate regardless of group. This can be especially problematic when particular subgroups of a population are underrepresented, leading any unique relationship effects to be suppressed in favour of the larger group effects. Ideally, researchers should be able to leverage the information found in large data sets for the good of all subgroups while simultaneously identifying key relationships that differ from one group to the next. To this end, a joint regularization method is proposed based on size-weighted joint regularization. The new method not only shares information across groups, but it also allows an identified group to differ in its modelling to ultimately result in both better prediction and estimation over the full data set.