Comparison of DIF Detection Performances of Mantel Test and Likelihood Ratio Test

摘要

The purpose of this study was to investigate Type I error rate of the IRT-Likelihood Ratio (IRT-LR) statistic and Mantel Test in detecting DIF. A multiple replication Monte Carlo study was utilized for simulated data sets. In final study design, there were 18 conditions [3 (sample size) × 3 (group mean difference) × 2 (methods of DIF detection)]. WinGen3 was used to simulate ability estimates and to generate response data sets. MULTILOG and DIFAS were used to conduct the Mantel and IRT-LR DIF analyses. Results indicated that with equal group distribution, Mantel Test and IRT-LR Test performed similarly under all testing conditions and had better Type I error rate control. Large sample size and presence of group mean difference tended to inflate the Type I error rates of both DIF detection tests. IRT-LR had higher Type I error rates than Mantel Test when large sample size and when group mean difference conditions.