The Need and Methods for Routine Inclusion of Model Uncertainty in Statistical Results

• Activity Number: 114 - The Need and Methods for Routine Inclusion of Model Uncertainty in Statistical Results
• Type: Topic Contributed
• Date/Time: Monday, July 31, 2017 : 8:30 AM to 10:20 PM
• Sponsor: Quality and Productivity Section
• Author(s): Hsin-wen Chang* and Hariharan Iyer and Yi-Ching Yao
• Companies: and National Institute of Standards and Technology and Institute of Statistical Science, Academia Sinica
• Keywords: Likelihood ratio ; Location-shift model ; Nonparametric statistics

Abstract

Data from multiple sources are frequently encountered in practice, and it is often of interest to combine information from these data to infer a common characteristic. This talk develops inference methods for such data, with different sources modeled by different shift parameters in a family of distributions. Nonparametric confidence band and hypothesis test are then developed for a reference distribution function of the family. The procedures are formulated using empirical likelihood, and calibrated by a bootstrap approach. Simulations are conducted to investigate robustness of the proposed methods to uncertainty in the specification of the distribution family. The method is illustrated using data from a forensic science example to address model uncertainty.