| This dissertation develops state space based EM algorithms for longitudinal mixed models. These models are exact for models with Gaussian errors, and include between subject random effects as well as within subject serial correlation for equally or unequally spaced data. The method uses the Kalman filter and smoothing algorithm to obtain the conditional estimates of the unobserved data given the observations. The methods are extended to non-Gaussian distributions from the exponential family providing approximate methods based on linearization techniques. Using simulated data, the method is compared with the new SAS macro GLIMMIX for Poisson distributed observations. |
