Estimating test score reliability when no examinee has taken the complete test

摘要

In some settings, the reliability of test scores must be approximated even though no examinee has taken the complete instrument. Rather, one group of individuals has taken one fraction of the items; a second independent group has taken a different, nonoverlapping fraction of the items; and so on. Ultimately, a representative sample of examinees will take the compete test, but at the time when the estimate of reliability is required, no such sample is available. Formulas are developed in this article to cope with this situation. The choice of one formula over another is contingent on the degree of similarity between or among the part-tests. Different estimators are developed for part-tests that are judged to be classically parallel, tau-equivalent, or congeneric. Standards are proposed for differentiating among these three models for the part-tests.