Better theory for better tests?

Abstract

So-called authentic language tests may not have recognizable separate items. Alternatively such tests may involve items which are dependent on each other, in a fixed order, and which cannot be replaced by alternatives. In neither case can classical test analysis be used at the item level. The use of the Rasch Model is also inappropriate. However, where total test score is the result of adding scores on a number of different, independent but equivalent parts, it is possible to estimate test reliability and to carry out item analysis on the basis of parts. In order to examine the homogeneity of test parts, the unidimensionality of the total score and the question of scale level, the CLA Model may be used. An example is provided of the use of the CLA Model in relation to a German C-Test.