Survival Analysis for Attrition Study and Informative Censoring I am trying to develop a model to understand attrition among students. Apart from binary classification, I am thinking of performing a survival analysis by following the students for 3 years at the end of which they will graduate. If I consider "attrition" as the "event" and the survival analysis would give me the chances that the student will continue their course for 3 years, the right-censoring of data is defined by an event which is "graduation".
Is survival analysis suitable for my attrition study ? If so, will the fact that the censoring of data is define by the fact the student has graduated affect the analysis. Further to this, there could be few cases where students take longer to graduate say for example 4 years.
 A: The problem with your proposed design is that it doesn't distinguish censoring due to graduation from censoring due to loss of follow up (e.g., student moves away and attends a different school, or student dies before graduation). That would seem to be an important distinction.
What you probably need is a multi-state model, including at least the following states: in school, dropped out before graduation, graduated. That would allow for someone to drop out, return later, and eventually graduate, while distinguishing graduation from censoring. If these are college-level or post-graduate students and you have the necessary data, you also might want to include a pre-matriculation state, modeling the transition from that state into the "in school" state. This vignette shows how to develop multi-state survival models in R.
One caution has to do with how frequently you are collecting data. In principle, students could drop out at any time during the year so if you have actual dates then a continuous-time survival model makes sense. If you are only collecting data at the ends of terms or academic years, however, you might be better off with a discrete-time survival model.
