Dynamic Models

Abstract

An approach to longitudinal data, similar in some ways both to autoregression and to random effects, and in fact able to encompasses both, involves allowing the regression coefficients to be random, evolving over time according to a Markov process. This is called a dynamic generalized linear model and is usually estimated by a procedure called the Kalman filter. (Unfortunately, the usual software for generalized linear models generally cannot easily be adapted for this algorithm.) Although originally proposed as the dynamic linear model for normal data, it can be extended to other distributions.