Abstract: This paper proposes a new method for stochastic analysis and control which does not require a model, but which is constructed directly from a raw database of patient responses to therapy. Roughly speaking, the basic idea is to evaluate a control (a therapeutic policy or modality) which has, on the average, proved to work well for similar patients in the database. By "similar" is meant patients who have the same covariates and who are in similar dynamical states. These concepts will be made more precise in the paper. The proposed stochastic analysis and control approach for databases is new, although it is motivated by methods of machine learning put forth in [1][2] and methods of dynamic programming for stochastic control given in [3][4].