The big data imposes the key problem of generalizability of the results. In the present contribution, we discuss statistical tools which can help to select variables adequate for target level of abstraction. We show that a model considered as over-fitted in one context can be accurate in another. We illustrate this notion with an example analysis experiment on the data from 13 university Massive Online Open Courses (MOOCs). We discuss statistical tools which can be helpful in the analysis of generalizability of MOOC models.