The effects of item position on item statistics were studied in a large set of data from tests of word knowledge (WK) and arithmetic reasoning (AR). Position effects on item response theory (IRT) parameter estimates and classical item statistics were also investigated. Data were collected as part of a project to refine the Army's Computerized Adaptive Screening Test (CAST). The CAST is an adaptively administered battery consisting of a WK subtest and an arithmetic reasoning (AR) subtest. As part of this effort to refine the CAST, 275 new and existing items from the WK and AR subtests were administered to 20,071 Army recruits from five different Army posts. A total of 270 of the items to be calibrated for each subtest was divided into six non-overlapping sets of 45 items each. The remaining five items were included in all six forms as potential anchors should subsequent equating prove necessary. Item statistics were computed separately for forward and reversed versions of each form. IRT parameters and classical parameters were determined. Estimates for both parameters varied significantly with item position. The variation was not generally related to the characteristics of the item, but was related to the ability of the examinees. There were no significant position effects when average percent passing scores were 75% or higher; position effects were quite pronounced when passing scores were 50% or lower. The primary conclusion drawn is the need to avoid unquestioning adoption of IRT methodology. Including the reversed version of each form prevented the introduction of systematic errors in the IRT parameter estimates. Seven data tables and 11 graphs illustrate the study findings. (SLD)