Abstract
Precise outcome predictions at an individual level from diverse genomic data is a problem of great interest as the focus on precision medicine grows. This typically requires estimation of subgroup-specific models which may differ in their mean and/or variance structure. Thus in order to accurately predict outcomes for new individuals, it’s necessary to map them to a subgroup from which the prediction can be derived. The situation becomes more interesting when some predictors are common across subgroups and others are not. We describe a series of statistical methodologies under two different scenarios that can provide this mapping, as well as combine information that can be shared across subgroups, with information that is subgroup-specific. We demonstrate that prediction errors can be markedly reduced as compared to not borrowing strength at all. We then apply the approaches in order to predict colon cancer survival from DNA methylation profiles that vary by age groups, and identify those significant methylation sites that are shared across the age groups and those that are age-specific.