The purpose of this paper is to prove that one currently recommended method of obtaining the reliability of an instrument defined on a population of aggregate units is invalid. This method randomly splits the aggregate into two halves, correlates the two half unit scores by a Pearson product moment correlation coefficient, and corrects the correlation coefficient using the Spearman-Brown prophecy formula. Our approach was to compare this procedure to the standard method of forming random split halves of items on the test. In addition the reliability of an instrument was obtained by both methods. It was found that the currently recommended method is an underestimate of the reliability of a test defined on an aggregate. (Author)