Content area
Abstract
Educational and psychological tests are critical for measuring latent traits, yet their fairness can be compromised by Differential Item Functioning (DIF), where individuals of similar abilities across demographic groups have unequal probabilities of correct responses. To address these challenges, this study introduces the Fair Rasch Model (FRM) and Generalized Fair Rasch Model (GFRM), which integrate fairness regularization into the Rasch model framework to mitigate DIF effects during parameter estimation without requiring prior DIF detection. These models use adjustable hyperparameters to balance fairness and estimation accuracy.
Simulation studies demonstrate that FRM and GFRM outperform existing methods in ability estimation, especially under conditions with high DIF magnitude or prevalence. In real data analysis using TIMSS 2015 mathematics assessments, the models minimized gender disparities in ability estimates more effectively than existing approaches. This study advances equitable testing practices, offering a novel approach to treat DIF in psychometric assessments.
