Longitudinal studies using a two-phase sampling design offer a great potential in identifying cases of rare chronic disease whose diagnosis is complex and expensive. In addition to ethical reasons, the detected cases provide precious resource for many other studies that require sufficient number of cases to make reliable conclusions. In this paper, we propose a formal statistical model to identify subjects at high risk of dementia based on the screening/diagnosis history in order to increase screening yield. We show that compared with cross-sectional method,up to 10% improvement in sensitivity can be achieved at several specificity levels of practical interest and the overall gain in prediction accuracy as measured by the AUC under the ROC curve is 3%-5%. Moreover, the inclusion of screening/diagnosis history is more helpful when the screening is less accurate. Our model provides a general framework that can be applied to many two-phase longitudinal studies of various diseases. It also serves as a preliminary investigation on an �adaptive� study design, in which probabilities of disease at each data collection wave are calculated based on the screening/diagnosis history and used to select subjects for diagnosis.

dementia, longitudinal, prediction, screening, two-phase.