查看文章

Understanding exam score distributions has implications for item response theory (IRT), grade curving, and downstream modeling tasks such as peer grading. Historically, grades have been assumed to be normally distributed, and to this day the normal is the ubiquitous choice for modeling exam scores. While this is a good assumption for tests comprised of equally-weighted dichotomous items, it breaks down on the highly polytomous domain of undergraduate-level exams. The logit-normal is a natural alternative because it is has a bounded range, can represent asymmetric distributions, and lines up with IRT models that perform logistic transformations on normally distributed abilities. To tackle this question, we analyze an anonymized dataset from Gradescope consisting of over 4000 highly polytomous undergraduate exams. We show that the logit-normal better models this data without having more parameters than the normal. In addition, we propose a new continuous polytomous IRT model that