Currently the high rates of university dropouts and low graduation are social problems that are very relevant in Chilean society. Predicting these events can allow institutions to take action to avoid them. The typical prediction models based on machine learning are capable of making reliable predictions, however they do not allow to understand the causality that originates both events, which could help to take better actions. This work proposes to find, analyze and weigh the causal relationships that allow predicting whether a student will drop out or will graduate according to the information available using a framework with Bayesian networks. The study is based on real data from the Universidad Católica de Temuco in Chile collected over three years. The results reveal variables and relevant relationships according the opinion of human experts, which suggest that the proposed model provides better capabilities to represent the causality of university dropout and graduation. From the results we believe that it is feasible to design better retention policies and timely degree at a university.