Detecting Differential Item Functioning From The Multi-process Approach

Abstract

Multi‐process IRT or (IR)tree models use a tree‐like approach to describe the process of reaching a response category in Likert‐type data and perform well in detecting extreme response style (ERS). These approaches usually divide the process into three steps: (1) indifference; (2) direction; and (3) intensity. A binary pseudo item (BPI) is created in each step, and such decomposition would inevitably result in nonimputable missingness in the second and third steps. Then, these BPIs are then examined with simplestructure multidimensional IRT models. Up to date, however, none of studies has investigated differential item functioning (DIF) under the framework of the tree approach. The present study examined how the logistic regression (LR) and the odds ratio (OR) methods performed in DIF detections through a series of simulations. Results showed that the OR method was rather robust across all DIF conditions and yielded satisfactory false positive rates (FPRs) and true positive rates (TPRs), whereas the LR method yielded inflated FPRs, especially when fitting to the BPIs derived from the second and third steps.