We describe three iterations of a Massive Open Online Course (MOOC) developed from online preparation materials for a reformed introductory physics classroom at the Massachusetts Institute of Technology, in which the teaching staff interact with small groups of students doing problems using an expert problem-solving pedagogy. The MOOC contains an e-text, simple checkpoint problems and homework. We show how certain course design aspects affect student behaviour: (a) frequent quizzes correlated with students reading a large fraction of the e-text, and (b) When homework sets are arranged by increasing (instructor-estimated) difficulty, we found strong correlations between difficulty and time to solution, but weak correlations with percent correct. Modifications to the second offering of the course resulted in higher retention. These modifications included targeting physics teachers and posting materials well in advance. We define retention as certificates earned relative to participants who make a significant effort on the second assignment. Retention measured this way varied between 44% and 72%, being highest in the course aimed at teachers. We show that there is significant learning among MOOC participants. Applying item response theory to common homework problems showed that the MOOC participants had significantly higher ability than students in a Massachusetts Institute of Technology course and that they maintained this advantage over the duration of the MOOC.