Toward More Robust Inferential Procedures for Coefficient Alpha Under Sampling of Both Subjects and Conditions

Barchard and Hakstian (1997b) demonstrated that inferential procedures used with coefficient alpha are not robust under sampling of both subjects and conditions (Type 12 sampling) with measurement data departing from essentially-parallel form. In the first of 2 studies reported here, a sample-based, non-analytical degrees-of-freedom correction factor was empirically developed that correlated almost perfectly with the independently-established correct degrees of freedom for the data. In the second study, application of this correction factor was assessed through a comprehensive simulation study involving Type 12 data sets representing a wide range of design characteristics and manifesting tau-equivalent measurement. Use of the correction factor yielded actual Type I error rates closer to nominal values than were obtained using uncorrected inferential procedures. Implications for practice and future research are discussed.