The popularity of online courses with open access and unlimited student participation (MOOCs) grows intensively. Certificates of MOOC completion are becoming a significant element in a student’s portfolio as well as an additional source of academic credits. That is why students, professors, universities, and employers have an interest in accurate measures of student’s proficiency in MOOCs. However, the current psychometric theories and non-psychometric approaches do not serve perfectly as measurement theories for MOOCs. At the same time, item response theory (IRT) seems to be a flexible and well elaborated framework that could be tuned up to be applied for MOOCs. Thus, the general purpose of the project is to propose extensions that make IRT able to accurately measure the student’s proficiency and to describe constructs closely linked with the student’s proficiency under the conditions of MOOCs.

The doctoral project consists of five studies. In the first study, we propose an IRT extension that measures the proficiency within a module in the case of an insufficient number of responses on test items and shows a growth of the proficiency between attempts to solve a certain item. In the second study, we propose an IRT extension that shows a growth of proficiency between modules. This extension uses student’s logs about each video lecture watched, each reading assignment viewed, and each item solved. In the third study, we apply an IRT extension to explore cheating in exams in MOOCs and to understand the student’s proficiency better. Using this extension, we study the differences between two groups of students – cheaters and non-cheaters. In the fourth study, we propose an IRT extension that predicts and explains the dropouts. We understand the student’s immunity to dropping out as a type of his/her proficiency. Finally, in the fifth study, we try to make the student’s proficiency parameter less biased. We explore one extra latent variable, the interest, and its influence on the student’s performance and the proficiency measures to separate the clean proficiency parameter. The five studies use (a) the common psychometric framework of cross-classification multilevel IRT models, and (b) real datasets from the Coursera platform.