A procedure is described for increasing the reliability of tests after they have been given and for developing shorter but more reliable tests. Eight tests administered to 200 graduate students studying educational research are analyzed. The analysis considers the original tests, the items loading on the first factor of the test, and the items which fit the Rasch model. Adjustments for test length are considered. Tests shortened by deleting items which do not fit the Rasch model have a higher internal consistency reliability then the longer original tests. This finding contradicts the theorem of classical test theory which states that increasing test length increases test reliability. The explanation for the find is that the theorem of classical test theory assumes that all items are homogeneous. Items not fitting the Rasch model are not unidimensional in the "latent-trait" sense which corresponds to being non-homogeneous in the "classical-test-theory" sense. Deleting such items results in a proportionately longer test in that a larger proportion of the remaining items is homogeneous. The misfitting items would not have been deleted on the basis of a classical item analysis. (Author/BW)