Combining Probabilities
Robert Rosenthal · 1991
approved by: reference: GAIA-C8-TN-MPIA-CBJ-053 issue: 2 revision: 1 date: 2011-12-20 status: Issued We show how to combine posterior probabilities from an ensemble of models, each of which estimates the same parameter (or class) but using “independent ” data. From this we describe how to separate out and replace the class prior (or the model-based prior) of a classifier post hoc and show how this relates to the combination problem. We also discuss the subtleties of conditional independence, what “independent data ” means, and outline under what circumstances dependent variables can become independent when conditioned on new information.