The effects of nonnormality and number of response categories on reliability

A computer simulation was designed to determine whether increases in reliability would be obtained for situations in which an observed categorical distribution, such as that for a Likert scale, was deliberately constructed to match the shape of the underlying distribution of the trait being measured. Results indicated that reliability did increase with the degree of similarity between the underlying and observed distributions. Reliability also increased with the level of correlation among the variables and with the number of scale points. However, maximum gains were reached with 5 or 7 scale points, after which reliability values leveled off.