Online Program Home
  My Program

Abstract Details

Activity Number: 529 - SPEED: Machine Learning
Type: Contributed
Date/Time: Wednesday, August 2, 2017 : 10:30 AM to 11:15 AM
Sponsor: Section on Statistical Learning and Data Science
Abstract #325270
Title: Estimation and Model Selection by Data-Driven Weighted Likelihoods
Author(s): Sam-Erik Walker* and Nils Lid Hjort
Companies: Dept. of Math., Univ. of Oslo and University of Oslo
Keywords: weighted Kullback-Leibler ; weighted likelihood ; data-driven weights ; robust inference ; robust model selection
Abstract:

We propose a weighted maximum likelihood methodology for parametric inference and model selection based on a weighted and penalized Kullback-Leibler distance function. Weighted likelihood is an estimation method worked with by several authors, from different perspectives, including Hjort (1994b), Hjort and Jones (1996), Loader (1996), Eguchi and Copas (1998), Claeskens & Hjort (2008) and Schweder & Hjort (2016). In these approaches, the weight-function is predefined and independent of the actual data. New in our current approach is that this weight-function can also be data-driven, i.e. determined from the same data as being used to estimate the model parameters and perform model selection. In principle, any data-driven non-negative weight function can be used, for example to bring extra estimation efforts into certain regions of the sample space, or to down-weight more extreme values for robustness purposes, or a combination of these. Theoretical properties of the resulting estimator and accompanying model selection criterion are given, together with several illustrations of the proposed methodology for simulated and real case examples of i.i.d. and regression type of data.


Authors who are presenting talks have a * after their name.

Back to the full JSM 2017 program

 
 
Copyright © American Statistical Association