Detection Power of Multilevel Latent-Trait Differential Person Functioning: A Monte Carlo Comparison with Conventional Person Misfit Statistics

Research on aberrant response patterns that deviate from modeled prediction has received considerable attention in the psychometric literature. A recent line of inquiry uses the multilevel modeling approach for the detection of differential person functioning in latent trait models in a hierarchical data structure. Preliminary Monte Carlo studies comparing the performance of this innovative approach with conventional person misfit statistics have yielded encouraging results. However, systematic evidence is lacking in demonstrating its power performance and robustness across misfit conditions. The present study extends previous research by simulating three different types of person misfit: spuriously high, spuriously low, and random guessing, with gender bias as the group-level effect. The proposed approach along with three other conventional indices was then applied to the simulated data to compare their relative performance in recovering those misfitting simulees and detecting gender bias. The entire experiment was replicated by Bootstrap sampling to evaluate its stability performance. Preliminary results suggest that power performance of the proposed method is either superior or comparable to that of the conventional indices. Suggestions for future research to manipulate potential confounding or moderating factors are discussed.