Dynamic Linear Model Diagnostics

Authors: 
Jeff Harrison, Mike West
University of Warwick, Duke University

Nov 30 1989

In time series analysis using dynamic linear models, retrospective analysis involves the calculation of filtered, or smoothed, distributions for state parameters in the past. we develop and illustrate novel results that are useful in retrospective assessment of the influence of individual observation on such distributions. In particular, new and computationally simple filtering equations are derived for past state parameters based on leaving out one observation at a time, providing dynamic model based versions of methods currently used in standard, static regression diagnostics.

Keywords: 

diagnostics, dynamic linear model, influence, jackknife calculations, outliers, smoothing time series

Manuscript: 

PDF icon 1990-09.pdf