Carnegie Mellon University

Causal Models from Online Course and Tutor Logs: Richard Scheines

Abstract: I discuss the sorts of log data available from online courses and tutors, the challenges that we face building meaningful variables from the raw log data, and the sorts of causal questions that provoke interest. Unlike many other domains, it is typical in educational contexts to have an exogenous, randomly assigned treatment variable as well as pre- and post-test measures that define "learning gain." Thus estimating the total effect of treatment on learning is not a search problem, but an estimation problem. Finding the mechanisms by which learning lakes place, however, is search problem. I discuss the general problem of mechanism discovery, and illustrate with one case from an online course and another from a cogntive tutor on fractions.