A Generalized Multi-Detectors Combination Approach for Differential Item
Functioning Detection
Abstract
To achieve more accurate and robust Differential Item Functioning (DIF)
detection, this study introduces a novel approach called Multi-Detectors
Combination (MDC). The main innovation of the MDC approach lies in
effectively integrating multiple existing single DIF detection methods
under observable examination conditions. Four specific supervised
learning methods (LR, NB, TAN, and SVM) in MDC are evaluated using
simulated data (360 testing conditions). The results indicate that MDC
methods outperform single DIF detection in accuracy (AUC: 81.2% to
82.3%), surpassing the AUC values of other methods, which ranged only
from 59.4% to 77.4%. Among MDC methods, TAN performed the best, under
all observable test conditions. Moreover, MDC methods demonstrate
enhanced precision (RMSE: 0.528 to 0.531) in ability estimation.
Analysis of actual data further validates the MDC approach’s superior
stability in ability estimation, achieved by excluding flagged DIF
items.