Improved Analytic Interval Estimation of Scale Reliability

Abstract

Since the beginning of the past century, the concept of reliability has attracted an impressive amount of interest by methodologists and substantive researchers across the social and behavioral sciences. The reliability coefficient represents an overall (unconditional) index of ‘precision’ of assessment, the proportion of true individual differences in observed variance. The attention it has received in the literature is appropriate, given the well-known fact that measurement in these disciplines is typically plagued by sizeable measurement error and related problems resulting from imperfect assessment. Over the past few decades, a number of methods have been proposed that aim at estimation of reliability, in particular of sum scores associated with multiple-component measuring instruments such as scales, questionnaires, tests, self-reports or inventories. These instruments are very often employed in social, behavioral and educational research, and a main reason for their popularity is that they provide multiple, converging pieces of information about underlying latent dimensions of major interest in these sciences, such as motivation, attitude, intelligence, social phobia, ability. For many years, Cronbach’s coefficient alpha (α) (Cronbach, 1951) has been a very frequently used index of reliability of multi-component instruments. At the same time, for quite a while its important limitations had not received the attention they deserve by empirical researchers.