In an example embodiment, submitted confidential data of a certain cohort (e.g., title, region, organization) is split into two components covering different portions of the cohort attributes (e.g., a first portion for (title, region)-wise confidential data values and a second portion of organization-wise compensation adjustments). These two portions are then analyzed separately and the inferences from both models are integrated together to obtain predictions for compensation values for the cohort as a whole.