A comparison of multivariate methods for the detection of Differential Item Functioning
Stephanie Hubert · Open access LMU (Ludwid Maxmilian's Universitat Munchen) · 2017
This thesis focuses on the comparison of three, recently developed, methods for the detection of differential item functioning (DIF) in the measurement of latent traits, such as abilities or attitudes in psychological or educational research.Identifying group differences is crucial for the correct and unbiased assessment of questionnaires.Over time, various methods have been proposed in literature to identify test items where DIF is present, which range from test statistics to modeling approaches.Most of these methods have some drawbacks in terms of usability or underlying assumptions, e.g. that they cannot deal with multi-categorical variables or that they focus on the global test level and do not identify DIF on the item level.The methods presented in this thesis, however, represent an advancement in the sense, that they try to overcome these problems and limitations.A commonality of the methods, that are described in the following, is, that they can cope with both multiple, potentially DIF-inducing, variables and any form of predictor variables, either metric or categorical.The advantage is a flexible and less restricted approach for the detection of DIF.The first considered method is called DIFlasso and is based on an extension of the widelyknown Rasch model, that involves additional group-specific parameters to incorporate group differences.DIF-detection is performed using a penalized estimation approach.The second method, DIFboost, uses boosting techniques to determine additional group-specific parameters in the extended Rasch model by means of iterative updating of so-called base learners.The third approach called DIFtree relies on model based recursive partitioning resulting in a decision tree for every item that carries out DIF.The aim of this thesis is to compare the three methods regarding their methodological approaches and by means of both an extensive simulation study and an applied example.