Abstract
Differential item functioning (DIF) is a phenomenon when two respondents with the same underlying latent knowledge but from different social group have different probabilities to answer item correctly. Many methods for DIF detection are derived from comparison of regression curves of reference and focal group.
Most of these approaches are limited in detection of DIF caused either by difference in difficulty or discrimination parameters with the exception of 3-4 parametric logistic item response theory (IRT) model (Birnbaum, 1968; Barton & Lord, 1981) and non-IRT models (Drabinová & Martinková, 2017). We introduce a novel approach using kernel smoothing estimation based on nearest neighbors based on general approach proposed by Srihera and Stute (2010).
We argue that this new approach has a great application potential, as it considers not only differences between groups caused by various difficulties and discriminations but also the differences in probability of guessing correct answer or in probability of inattention when answering.