Overfactoring in rating scale data: A comparison between factor analysis and item response theory.

Abstract

Educational and psychological measurement is typically based on dichotomous variables or rating scales comprising a few ordered categories. When the mean of the observed responses approaches the upper or the lower bound of the scale, the distribution of the data becomes skewed and, if a categorical factor model holds in the population, the Pearson correlation between variables is attenuated. The consequence of this correlation attenuation is that the traditional linear factor model renders an excessive number of factors. This article presents the results of a simulation study investigating the problem of overfactoring and some solutions. We compare five widely known approaches: (1) The maximum-likelihood factor analysis (FA) model for normal data, (2) the categorical factor analysis (FAC) model based on polychoric correlations and maximum likelihood (ML) estimation, (3) the FAC model estimated using a weighted least squares algorithm, (4) the mean corrected chi-square statistic by Satorra-Bentler to handle the lack of normality, and (5) the Samejima's graded response model (GRM) from item response theory (IRT). Likelihood-ratio chi-square, parallel analysis (PA), and categorical parallel analysis (CPA) are used as goodness-of-fit criteria to estimate the number of factors in the simulation study. Our results indicate that the maximum-likelihood estimation led to overfactoring in the presence of skewed variables both for the linear and categorical factor model. The Satorra-Bentler and GRM constitute the most reliable alternatives to estimate the number of factors.