I'm using logistic regression to predict student retention in an online course.

I have a data of student interactions within a web platform of an online course. The course spans 6 weeks, with new lecture resources and new assignments uploaded at the beginning of each week. The weekly assignments due at the end of each week. Students can watch lecture videos, view/write forum posts, and do assignments. Students can drop out of the course any time (i.e. no longer interacting within the course platform).

Each week, I want to predict the likelihood of a student staying within the course in the next week (stay in the next week=1, out in the next week=0). The predictors are the number of times the student watch the lecture videos (video_views), the number of posts the student read (posts_read), the number of posts the student wrote (post_written) and the student's score of this week assignment (score).

I'm thinking of building 6 models using logistic regression, for each week. But I also want to make a connection between, say, week 6's model and week 5's model. Can such connection be shown if I use percentage of cumulative values for each predictors (e.g. cumulative number of assignment score out of total assignment score for the whole course) instead of weekly values?

Should I just build a single model with the course week (course_week) as another predictor? This is my first time using logistic regression, and I'm not sure about putting a time variable in a logistic regression model.

  • 2
    $\begingroup$ What you are doing, in essence, is discrete time survival analysis. Here is some information from UCLA's stats help site. I haven't seen these slideshows, but the UCLA site is of uniformly high-quality. $\endgroup$ – gung Oct 21 '13 at 2:54
  • $\begingroup$ Thanks for the pointer, gung. I've done a quick read on the UCLA's slides, and it seems their model measures the survival probability of a reference subject at any time based on several preconditions of that subject and an early observation. $\endgroup$ – Agus Dwikarna Oct 22 '13 at 21:24
  • $\begingroup$ The subjects in my model have no preconditions data (we don't have access to students' data other than students' interaction within the course), and the observations are happening weekly. Am I missing something? Thanks in advance. $\endgroup$ – Agus Dwikarna Oct 22 '13 at 21:34
  • $\begingroup$ Couldn't you use proportional odd logistic regression? Outcome is number of weeks "stayed in". $\endgroup$ – DWin Nov 2 '13 at 2:39

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.