Revisiting Reliability: The Misuse of Cronbach's in Political Science.

As the recent exchange in Psychometrika illustrates (Green and Yang 2009; Sijtsma 2009), 60 years after the original paper (Cronbach 1951), the psychometric community is still trying to instruct readers on how to meaningfully interpret and report Cronbach's α. Most students of public opinion know that α is not the most sophisticated measure of a scale's internal consistency (Borsboom 2006; Bentler 2009). However, it remains the most often cited measure of reliability. More disturbingly, most researchers use a "large enough" coefficient α to illustrate something that is not mathematically warranted (unidimensionality or trait strength). Given the slippage between advanced statistical methods and those who use reliability analysis in applied public opinion research, this paper sets out to illustrate a new way to think about Cronbach's α and to show how the size, and therefore "statistical significance," of α is largely a function of the number of items in the analysis and the number of subjects taking the test. Through Monte Carlo simulation, this paper simplifies interpretations of α by clearly deriving the limits of its usefulness, illustrating its sensitivity to sample and scale sizes, and providing users with an additional measure to make α comparable across different scales and samples. [ABSTRACT FROM AUTHOR]