Retrieving IRT parameters with half information

Marie Sacksick (1, 2) , Sebastián Ventura

1 Universidad de Córdoba = University of Córdoba [Córdoba]
2 CHART - Cognitions Humaine et ARTificielle

Item Response Theory can be used to estimate the degree of mastery of a concept by learners, to automatically assess their knowledge. The models stemming from this theory are tuned to be adapted to the questions used to assess mastery. The correct estimation of the parameters is key to be able to have a correct estimation of the mastery. However, this estimation can be skewed by missing data, noise on the model, or a lack of data. The question we ask here in this paper is how much data, created by a given number of students answering to a given number of questions is necessary to retrieve reliable coefficients of the questions, when the database at disposal have missing data. To do so we use simulated data. There are two case studies with different levels of data emptiness: one is the baseline and has complete information, the other has only half information. We find that even though IRT models seem robust against missing values, it is not possible to use the thresholds of the literature obtained with a full database.

Mots clés: item response theory; parameter estimation; missing values

Domaines: Machine Learning [stat.ML]; Psychologie

Marie Sacksick, Sebastián Ventura. Retrieving IRT parameters with half information. 11th International Conference on Educational Data Mining, Jul 2018, Buffalo, United States. ⟨hal-02263802⟩