Abstract
In many practical problems, there is interest in the estimation of mixed effect projections for new data that are outside the range of the training data. Examples include predicting extreme small area means for rare populations or making treatment decisions for patients who do not fit typical risk profiles. Standard methods have long been known to struggle with such problems since the training data may not provide enough information about potential model changes for these new data values (extrapolation bias). We propose a new framework called Prediction Using Random-effect Extrapolation (PURE) which involves constructing a generalized independent variable hull (gIVH) to isolate a minority training set which is "close" to the prediction space, followed by a regrouping of the minority data according to the response variable which results in a new (but misspecified) random effect distribution. This misspecification reflects "extrapolated random effects" which prove vital to capture information that is needed for accurate model projections. Projections are then made using classified mixed model prediction (CMMP) (?) with the regrouped minority data. Comprehensive simulation studies and analysis of data from the National Longitudinal Mortality Study (NLMS) demonstrate superior predictive performance in these very challenging paradigms. An asymptotic analysis reveals why PURE results in more accurate projections. for this article are available online.