Negative Effect of Item Heterogeneity on Split-Half Reliability Estimates: An Empirical Demonstration

Abstract

Two groups of items were deliberately selected from a standard psychometrically sound test so that both had the same mean value for ail the interitem correlations yet they differed in terms of the variance among the interitem correlations. The two groups were then treated as separate tests, and all possible split-half correlations were computed for each “test.” The more heterogeneous group resulted in a wider spread of split-half correlations. Furthermore, both groups of correlations showed a significant skew in the negative direction. It was concluded that an odd-even split and coefficient alpha significantly underestimate a test's true split-half reliability, as defined by the reliability coefficient generated by using the two most analogous forms of a test.