We study academic integrity in a final exam of a game theory course with 463 undergraduate students at a major Spanish university. The exam is an unproctored online multiple-choice exam without backtracking. A key characteristic is that for each (type of) problem, groups of students receive different versions. Moreover, each problem version is assigned to one subgroup during one stage of the exam and to another subgroup during an immediately consecutive later stage. Thus, we can exploit grade points and timestamps to study students' academic integrity. We observe a significant decrease in completion time at each later stage; however, surprisingly, there is no corresponding impact on average grade points. The precise number of different versions does not seem to have an effect on either variable. Our findings thus suggest that employing a limited number of distinct problem versions (as few as two) can diminish cheating effectiveness in online exams.
