| Abstract: | It is sometimes sensible to think of the fundamental unit of test construction as being larger than an individual item. This unit, dubbed the testlet, must pass muster in the same way that items do. One criterion of a good item is the absence of DIF–the item must function in the same way in all important subpopulations of examinees. In this article, we define what we mean by testlet DIF and provide a statistical methodology to detect it. This methodology parallels the IRT-based likelihood ratio procedures explored previously by Thissen, Steinberg, and Wainer (1988, in press). We illustrate this methodology with analyses of data from a testlet-based experimental version of the Scholastic Aptitude Test (SAT). |
