Answers are grounded in this paper and cite their sources.
Incomplete and mistimed longitudinal data are common problems in longitudinal studies of free living populations. One possible approach to the analysis of longitudinal data with randomly missing or mistimed data uses a mixed effects model. While the theory underlying this approach is not new (Harville, 1977), very little has been published about its actual application. The objective of this paper is to examine the behavior of the maximum likelihood estimates of mixed effects model parameters when the distribution of the parameters underlying the data is known. Specifically, the effects of different values of the error covariance parameters and missing and mistimed data are considered. Two iterative methods, the Method of Scoring and the EM algorithm, were utilized to solve the maximum likelihood equations. In the setting considered here, the Method of Scoring and the EM algorithm worked well for the analysis of longitudinal data with missing and/or mistimed data. The variance of the parameter estimates increased when the data were mistimed and/or missing. The Method of Scoring was computationally more efficient than the EM algorithm. Both methods converged to the same set of parameter estimates except when the covariance matrix of the random effects, D, was nearly singular. When this occurred the Method of Scoring sometimes produced non-positive definite estimates of D, The EM algorithm, while constraining [Dcirc] to be positive definite, often failed to meet the convergence criteria in the allotted number of iterations. Possible “solutions” to this problem are discussed.
No takes yet. Share an insight, caveat, or question.
Fairclough et al. (1986) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: