Files
Abstract
Multidimensional item response theory (MIRT) model selection has only recently come about following the development of the Mr -family of statistics (Maydeu-Olivares & Joe, 2005; 2006). Global model fit indices are now available for a wide range of MIRT models and are theoretically equivalent to typical χ2-based fit indices (e.g. RMSEA, SRMSR, CFI, and TLI). These fit indices are evaluated relative to popular cut-offs, such as those derived by Hu and Bentler (1998, 1999). The purpose of the present study was to establish whether these popular cut-offs achieve acceptable levels of Type 1 error for model selection or have adequate Power for detecting misfit in MIRT models. Results from two simulation studies suggest the performance of these cut-offs varies depending on the MIRT model of interest and study design characteristics and the power to detect model misspecification is contingent on the type of misspecification present.