West Coast Ballroom
WITHDRAWN - An Informative Stacking of Learner with Disjoint Predictor (307874)
David Nelson, CCDOR/University of Minnesot*Siamak Noorbaloochi, CCDOR/University of Minnesota
Khalil Shafie, University of Northern Colorado
Michele Spoont, CCDOR/University of Minnesota
Keywords: Variance of Likelihood rattios, resilency, Predictive Modelling
Assume we have conducted two distinct studies to predict resiliency as a categorical attribute. Each study, using alternative supervised methods provided a series of class memberships. Within each study the predicted class memberships can act as new predictors of a super learner to produce two stacked classifies. However, the two studies have two sets of possibly distinct predictors. Some of the predictors are common between the two and some are not, and hence, simple augmentation of the two data sets will result in artificial missingness. Using Kagan divergence (Variance of Likelihood Ratios), we propose a method to combine predictions from several studies with disjoint sets of predictors. The final classifier properties will be investigated, and the methodology will be applied to two mental health studies, where resiliency class membership was predicted with non-overlapping set of predictors. The extension of the methodology as a meta--analytic method for combining several health studies on continuous outcome predictions will be discussed.