Title: Bayesian prediction based on profile-reference data
Authors: Shigetoshi Hosaka - Hosaka Clinic of Internal Medicine (Japan)
Jinfang Wang - Chiba University (Japan) [presenting]
Abstract: The problem of predicting the future outcome for a specific subject is considered based on data resulting from a longitudinal study. This data will be referred to as the profile data, which typically contain the time-dependent responses for each subject, as well as many other variables concerning the background information on each subject. In addition to the profile data, we assume that there are also available a large reference data set containing the same time-dependent responses as in the profile data. The reference data may be obtained from national survey offices, which publish many kinds of survey data, such as data on health care. The reference data differ from the profile data in that individual data in the reference data are usually grouped and only very limited background information (e.g. gender and age) are available. We assume that the two data sets partially share a common latent structure. We propose a Bayesian model for predicting the future outcome for a specific subject by combining the profile and the reference data. We illustrate this methodology by applying a dynamic linear mixed model to predict the blood sugar level at a specific age.