We use administrative, claims, and clinical data from a major academic health center (AHC) to investigate (1) members’ health states and health state trajectories, and (2) optimal interventions to improve population health at more affordable costs. We use a simulation-based approach to predict intervention effects and their uncertainty in the AHC system. Our prediction of intervention effects (PIE) model is composed of six constituent models corresponding to member health state; probability of positive expenditure; size of positive expenditure; and disenrollment due to death, changing plan, or other reasons. Our health state model uses latent variables to identify groups of diagnostic codes that commonly co-occur. We then allow each individual’s class membership probability distribution to change smoothly through time as their underlying health state improves or worsens. A strength of our PIE model is its flexibility; the constituent models can be adapted to the scientific question of interest without changing the overarching structure of the system. The methods are illustrated with a study of diabetes and COPD in the AHC clinical population.