Outlier Detection in Dynamic Functional Models

• Activity Number: 329 - Novel Developments in Functional Data Analysis
• Type: Topic Contributed
• Date/Time: Tuesday, July 31, 2018 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Learning and Data Science
• Author(s): Andrada E Ivanescu* and William Checkley and Ciprian Crainiceanu
• Companies: Montclair State University and Johns Hopkins University and Johns Hopkins University
• Keywords: functional data analysis; dynamic prediction; dynamic outliers; function-on-function regression; longitudinal data; child growth

Abstract

We propose methods for dynamic identification of outliers for sparse or dense longitudinal data. We call these methods dynamic because the associated models can be applied at and tailored to any time point in the history of the data for one individual. Dynamic methods are different from static approaches that use all available data after it is collected. Static approaches are useful in retrospective studies when one is interested in data quality control, whereas dynamic approaches are useful when one is interested in identifying unusual observations as soon as possible and use these findings for interventions as data are acquired. The methods we propose can use covariate adjustment both for time-dependent and time-independent covariates. Methods are motivated by and applied to a child growth study. Using these data, we show that there is negligible overlap between the top outliers identified by our new dynamic models and static outlier identification methods.