Manual,
Cluster-based Anchor Item Identification and Selection (NEPS Working Paper No. 68)
Leibniz-Institut für Bildungsverläufe, Nationales Bildungspanel, Bamberg, Deutschland, (2017)
Abstract
In order to compare scores of latent variables across groups or measurement ocassions, the respective items presented to both groups or at both measurement occasions need to be measured invariant, that is, show no differential item functioning (DIF). In situations where this assumption is violated, researcher may strive for partial measurement invariance by identifying a set of items (anchor items) that are DIF-free. Different approaches for detecting DIF-free items exist. These either make the assumption of unbalanced DIF or the assumption that the majority of items is DIF-free. Recently, Bechger and Maris (2015) proposed an approach that instead of identifying DIF-free items identifies clusters of items that function similarly. As such they do not make the assumption of unbalanced DIF or that the majority of items is DIF-free. While this approach is very promising, it is not applicable, yet, for substantive research. 1. There is no clear criterion for the identification of clusters. 2. There are no criteria for choosing a cluster as anchor for linking purposes. (a) We propose two procedures for cluster identification, that are, the k-means clustering approach and the range-and-step-threshold approach. (b) We provide three selection criteria (cluster homogeneity, cluster accuracy, and cluster size) that may aid the choice of a cluster. For illustration, we apply the approach on data of a linking study in the National Educational Panel Study comparing reading competence between grade 9 students and adults. The paper closes with a discussion of the advantages as well as the limitations of the proposed methods and a delineation of further research areas.
