Coefficient alpha is a widely used and useful index of reliability, but it provides an inappropriate estimate of reliability when used with multidimensional composites. The problematic nature of this inappropriate usage varies primarily as a function of the prominence of distinct factors in a composite and to a lesser extent as a function of the length of the measure. Implications for use of coefficient alpha estimates of reliability as the basis for correcting observed correlations for unreliability are described for a variety of levels of multidimensionality and measure length. Researchers should correct for unreliability to estimate construct interrelationships but must exercise care as to how they conceptualize errors of measurement and how they estimate reliability.
Rogers, William M.; Schmitt, Neal; and Mullins, Morell E., "Correction for Unreliability of Multifactor Measures: Comparison of Alpha and Parallel Forms Approaches" (2002). Peer Reviewed Articles. 45.