In short, we want educational predictions to be wrong. If our predictive model can tell that a student is going to fail, we want that to be true only in the absence of intervention. If the student does in fact fail, that should be seen as a failure of the system. A predictive model should be part of a prediction-and-response system that (1) makes predictions that would be accurate in the absence of a response and (2) enables a response that renders the prediction incorrect (e.g., to accurately predict that, given a specific intervention, the student will succeed). In a good prediction-and-response system, all predictions would ultimately be negatively biased. The best way to empirically demonstrate this is to exploit random variation in the assignment of the system—for example, random assignment of the prediction-and-response system to some students but not all. This approach is rarely used in residential higher education but is newly enabled by digital data.The grand challenge in data-intensive research and analysis in higher education is to find the means to extract knowledge from the extremely rich data sets being generated today and to distill this into usable information for students, instructors, and the public.
Via
Peter Mellow