The Effects of Heterogeneous Item Distributions on Reliability

A computer simulation was designed to investigate the degree to which coefficient alpha is affected by the inclusion of items with different distribution shapes within a unidimensional scale. Theoretical scales were generated that combined different numbers of normal, moderately skewed, extremely skewed, and leptokurtic items. This process was repeated using different numbers of response categories (3, 5, 7, and 9), and different levels of interitem correlation (.25, .50, and .75). Results indicated that reliability did not decrease dramatically as a result of utilizing differentially shaped items within a scale; the largest decreases came about when utilizing highly skewed items. However, these results were moderated by the correlational structure and by the number of response categories. Utilizing differentially shaped items brought about the largest decreases in reliability when the interitem correlations were low and when the number of response categories was small. Conversely, reliability was least affected when the interitem correlations were high and when the number of response categories was large. It was suggested that the practice of deleting items from a scale on the basis of distributional nonnormality may adversely affect scale validity with no appreciable increase in reliability.