PAGE. Abstracts of the Annual Meeting of the Population Approach Group in Europe.
PAGE 25 (2016) Abstr 3687 [www.page-meeting.org/?abstract=3687]
Oral: Drug/Disease modelling
Marc Vandemeulebroecke, Johanna Mielke, Peter Quarg, Tillmann Krahnke, Bj÷rn Bornkamp, Andreas Monsch
Novartis Pharma AG; Memory Clinic Basel
Objectives: The goals of this work were to investigate which cognitive domains carry most information on earliest signs of cognitive decline in the elderly, to characterize the subjects' cognitive trajectory over time and to understand which subject characteristics impact this trajectory. This is relevant for better understanding whom to treat and what to measure in early intervention trials in slowly progressing neurodegenerative diseases such as Alzheimer’s disease (AD).
Methods: A longitudinal Item Response Theory (IRT) model was developed for cognitive data from the BASEL study, in which 1750 mostly healthy elderly subjects were observed over up to 14 years per subject. The model extends an earlier cross-sectional model  (which was inspired by ) into a fully longitudinal IRT model, in which the multifaceted nature of the response and its longitudinal trajectory are modeled jointly. It was implemented in a Bayesian framework with noninformative priors, using WinBUGS, JAGS and STAN.
Results: 'CVLT-Word List Learning' and 'CERAD-Word List Learning' as well as 'CVLT-Word List Long Delay Free Recall' and 'CERAD-Word List Delayed Recall' carried most information in the BASEL sample (15.5%, 13.1%, 10.3% and 8.8%, respectively, of the total amount of information). The Mini Mental Status Examination (MMSE) and word list recognition tasks were informative only in the range of low cognitive abilities. Greater age at baseline, positive APOE4 carrier status, and less years of education were significantly associated with a faster cognitive decline. WinBugs, JAGS and STAN provided virtually identical results. JAGS provided the best compromise between efficiency and practicality.
Conclusions: Fully longitudinal IRT modeling, as applied here in a mostly healthy elderly population, is a suitable method to capture the multifaceted nature of cognition and its longitudinal trajectory jointly. It is computationally more intensive than cross-sectional IRT models (such as  and ), but it allows the estimation of the IRT parameters based on all data. It would be of interest to apply this method also to a cohort with prodromal or mild AD.