Doing More with Less: Statistical Methods for High Dimensional and Correlated Regression Problems

Doing More with Less: Statistical Methods for High Dimensional and Correlated Regression Problems

Abstract

Modern statistical methods are spectacularly rich and complex. They have the potential to help practitioners answer difficult but important questions about fundamental scientific problems. However, there are often gaps between the how modern statistical methods are introduced in the statistical literature and their implementation, which can make them impractical or inefficient from the perspective of practitioners. This talk presents several specific examples of how existing modern statistical methods can be made more accessible. They include faster, less computationally demanding methods for fitting popular models, new generalizations of popular models with desirable properties, and ways to avoid unnecessary and limiting assumptions. They are motivated by challenges practitioners encounter when analyzing high dimensional data and correlated data using penalized or Bayesian regression models, long memory time series models, Log-Gaussian Cox process models, and nonlinear potentially black-box machine learning models.

Bio

Maryclare Griffin is an Assistant Professor in the Department of Mathematics and Statistics at the University of Massachusetts Amherst. Her research focuses on computational problems that arise in Bayesian statistics, penalized regression, spatial statistics, and time series.