A reassessment of graduation modeling for policy design

A vast and diverse literature estimates graduation chances using logistic models set in an arbitrary timeframe, where a graduation indicator is checked at a conventional point in time and associated with covariates measured at some date. Survival models emerged over time as a robust alternative, for being able to estimate time-to-degree and time-varying effects of predictors. This paper reconsiders the effectiveness of both modeling approaches in addressing policy-relevant questions, particularly in light of the increasingly automated and algorithm-based educational policies. We find that both methods exhibit blind spots and limitations, but that adopting a simple pragmatic approach logistic models can achieve a comparable level of effectiveness at depicting graduation dynamics while also being capable of answering questions that are problematic for survival models. We exploit a unique dataset and the nature of discrete-time survival models as combinations of logistic regressions run at different times to illustrate how arbitrary timeframes impact the estimates of a logistic model of graduation. Conversely, we illustrate how separately running and analyzing all the distinct logistic regressions provides insights that are unlikely to come from a survival model.