When Cronbach (1951) introduced coefficient alpha more than half a century ago, he correctly anticipated the usefulness of this index of internal consistency. It has proven to be the most popular reliability estimation method by far (Hogan, Benjamin & Brezinzki, 2000), and social sciences citations of Cronbach's 1951 article run into several hundreds per year (Cronbach, 2004). Cortina (1993) refers to this coefficient's use not only in the various areas of psychology, but also in sociology, statistics, medicine, counseling, nursing, economics, political science, criminology, gerontology, broadcasting, anthropology and accounting. Undoubtedly this coefficient's popularity may be attributed to the fact that it doesn't require more than one test administration (as does the test-retest method), or more than one parallel test form (as does the parallel-forms method), or the splitting of a test into two parallel halves (as do the split-half methods).Not only is Cronbach's 1951 article the source of the highly popular coefficient alpha, but it still continues to stimulate all kinds of methodological comments and developments. There has been a continuous stream of attempts to derive alternative coefficients for situations where the assumptions underlying coefficient alpha have been relaxed (cf. Lucke, 2005). Other methodological contributions (e.g., those of Cortina, 1993;Henson, 2001;Schmitt, 1996;Streiner, 2003 andThompson, 2003) have had a more modest aim, namely, to explain some of the anomalies and misconceptions associated with coefficient alpha. These contributions addressed, amongst others, the possibility of relatively high alpha values for multidimensional item data and the occurrence of negative alpha values. However, it would appear that these messages still haven't reached all of those who have access to computer programmes for the computation of coefficient alpha -it is not uncommon to find applications of coefficient alpha as a test of unidimensionality or as an internal-consistency index across subtests (of relatively independent constructs) in some locally published research.To explain the intricacies of coefficient alpha, the authors referred to above used numerical examples of item variancecovariance (VCV) matrices which may be less than helpful in explaining how such aberrant matrices have come about in the first place. The purpose of the present article is to revisit these earlier contributions but to start off the explanations involved in terms of simple item data matrices rather than item VCV matrices. In the process, some inconsistencies and omissions in at least some of these contributions will be noted and clarified. A clear understanding of coefficient alpha may prevent its incorrect use and interpretation. Finally, some of the alternatives to coefficient alpha for situations in which some of its assumptions do not hold will be mentioned. But first, a review of the assumptions, derivation and computation of this coefficient is presented.
ASSUMPTIONS AND DERIVATION OF COEFFICIENT ALPHAThe comp...