The robustness of confidence intervals for coefficient alpha under violation of the assumption of essential parallelism

Two methods exist for setting confidence intervals for coefficient alpha: Feldt's (1965) exact F-distribution method and Hakstian and Whalen's (1976) asymptotic normal-theory procedure. These methods both assume that items are essentially parallel: Two studies, employing Type 12 sampling, are reported in which the effects of violating this assumption are examined. The results indicate that as long as the data manifest the properties of essential parallelism, the two methods maintain precise Type I error control. With a variety of forms of measurement data containing correlated errors or unequal variances, however, the methods resulted in poor Type 1 error control. No performance differences between the two methods were found. Implications are noted for practice and for future research on inference with coefficient alpha.