Using innovative outliers to detect discrete shifts in dynamics in group-based state-space models

Outliers are typically regarded as data anomalies that should be discarded. However, dynamic or "innovative" outliers can be appropriately utilized to capture unusual but substantively meaningful shifts in a system's dynamics. We extend De Jong and Penzer's 1998 approach for representing outliers in single-subject state-space models to a group-based framework. The group-based model enables model predictions concerning the group as a whole while individualized change points are incorporated to capture discrete shifts in each individual's trajectory. An illustrative example using cognitive performance data from N = 38 older adults over 60 days is presented. Results indicated that a subset of older adults showed sudden shifts in learning dynamics in several forms, including sudden jumps in level, slope, and linkage between perceived control and actual performance level. The broader implications of incorporating innovative outliers into studies of intraindividual change are discussed.