|
Abstract:
|
The world of higher education offers important and impactful opportunities to apply predictive statistical models, including efforts to improve student success and retention, inform recruitment and admissions efforts, and more reliably support institutional budgeting and planning goals. Nevertheless, the annual academic cycle imposes challenges to using off-the-shelf statistical learning techniques. Not only does the true data generating mechanism change over the course of an admissions season (failure to accept admission means different things six months from the start of classes than it does one week prior), but the internal and external environments are constantly shifting; while we can collect nuanced student information, process, practice, and environment changes conspire to introduce difficult to detect sources of extrapolation error. In this work, we present both heuristic approaches to minimizing this unmodeled extrapolation risk, and a modified classification tree structure which uses a penalty to explicitly minimize a subset of such errors, focusing on the application of these techniques for undergraduate retention modeling at the University of Iowa.
|
