A Simple and General Changepoint Detector

• Activity Number: 222
• Type: Contributed
• Date/Time: Tuesday, August 10, 2004 : 10:30 AM to 12:20 PM
• Sponsor: Biometrics Section
• Author(s): Robert H. Riffenburgh*+ and Kevin M. Cummins
• Companies: Naval Medical Center San Diego and San Diego State University
• Keywords: changepoint ; time series ; longitudinal data ; Moving F ; CUSUM/EWMA ; process control

Abstract

A Moving F statistic to detect and locate changes in longitudinal data (time series) is easier to use than other commonly employed change point detectors (process control, CUSUM, EWMA, data-based bandwidth selection) and is shown to work in several situations where these other detectors fail or lack methodological development. In not-necessarily-stationary series, the expected mean path may take on a variety of functional forms (constant, linear, quadratic, exponential, etc.), in which any one or any combination of parameters may change (jump, slope, curvature, etc.), and the residual variability about the path may also change (increase or decrease). Furthermore, a decrease in variability may mask a change in a parameter. The Moving F statistic detects a change in any parameter or simultaneous set of parameters and/or variability, and also uncovers a masked change. Patterns in the forces giving rise to the data may often be perceived. An example from monitored PSA level in a patient equivocal for prostate cancer shows that cancer could have been detected nearly two years earlier than by conventional means.