Answers are grounded in this paper and cite their sources.
The traditional KR20 and KR21 estimates of reliability are well known for dichomotous data. This paper derives forms of KR20 and KR21 for polytomous data, which can be applied to Likert scales. This is accomplished by deriving the theoretical population reliability for polytomous data which follows the multinomial-Dirichlet model, where responses to each item are multinomial and response category probabilities for each individual follow a Dirichlet distribution. The KR20 and KR21 formulas for polytomous data are constructed as consistent estimators of reliability for this model, and are given in Equations (10) and (11). It is shown that the KR20 for Likert scale data is a linear transformation of Cronbach's alpha. In the process, a method for simulation of unidimensional Likert scale with a known population reliability is developed using the multinomial-Dirichlet model. Using the simulation method, a study compares the KR20 estimator of reliability to several estimators of coefficient omega and shows the relative strengths of KR20 as compared to omega: it performs nearly as well for large sample sizes and population reliabilities and is more numerically stable than the estimators based on a factor analysis for small sample sizes and population reliabilities. Furthermore, overfitting a factor analysis model can have serious negatively effects on the performance of reliability estimators based on the model. Lastly, future theoretical development for reliability in multivariate exponential family data are proposed.
No takes yet. Share an insight, caveat, or question.
Robert C. Foster (2022) studied this question.