Outlier Removal Using the Log-Likelihood for Group-Based Trajectory Modeling

• Activity Number: 498
• Type: Contributed
• Date/Time: Wednesday, August 6, 2014 : 10:30 AM to 12:20 PM
• Sponsor: Biometrics Section
• Author(s): Christopher Davies*+ and Gary Glonek and Lynne Giles
• Companies: University of Adelaide and University of Adelaide and University of Adelaide
• Keywords: Outlier removal ; Group-based trajectory modelling ; Longitudinal studies

Abstract

Attributes measured longitudinally can be used to define trajectories for each individual in a given population. Group-based trajectory modelling methods identify groups of trajectories such that trajectories within groups are more similar than trajectories in distinct groups. Existing methods generally allocate all trajectories into groups, however this overlooks the prospect some trajectories may be so different from the rest of the population that they should not be included in a group-based trajectory model. These outlying trajectories are then treated as though they belong to one of the groups, distorting the estimated trajectory groups and any subsequent analyses that use them.

We have developed an algorithm for removing outlying trajectories based on comparing the log-likelihood contributions of the observations to those of simulated samples from the estimated group-based trajectory model. At each step of the algorithm, the type of model and number of groups are chosen according to the maximum Bayesian information criterion, and then the outlying trajectories are removed. In this talk the algorithm will be detailed and an application of its use will be demonstrated.