The rationale underlying factor analysis applies to continuous and categorical variables alike; however, the models and estimation methods for continuous (i.e., interval or ratio scale) data are not appropriate for item-level data that are categorical in nature. The authors provide a targeted review and synthesis of the item factor analysis (IFA) estimation literature for ordered-categorical data (e.g., Likert-type response scales) with specific attention paid to the problems of estimating models with many items and many factors. Popular IFA models and estimation methods found in the structural equation modeling and item response theory literatures are presented. Following this presentation, recent developments in the estimation of IFA parameters (e.g., Markov chain Monte Carlo) are discussed. The authors conclude with considerations for future research on IFA, simulated examples, and advice for applied researchers.
Keywordsitem factor analysis; parameter estimation; categorical data; confirmatory factor analysis; item response theory Item-level data within the social and behavioral sciences are often categorical in nature. Dichotomous (e.g., disagree vs. agree) or polytomous (e.g., strongly disagree, disagree, neither, agree, and strongly agree) item response formats may fail to maintain the scale and distributional properties assumed by models such as ordinary least squares regression or common linear factor analysis. Traditional regression techniques describe the outcome variable as an optimal linear function of observed predictors. The proper implementation of these techniques requires assumptions such as independent and normally distributed residuals, a continuous conditionally normal outcome, and that the model is correctly specified (i.e., a linear relationship exists between the outcome and predictors). The common linear factor model, which describes the covariances among observed variables as a function of a smaller number of latent factors, makes many of the same assumptions. It is assumed that the unique factors (those that affect only one measured variable) are normally distributed, the outcomes are continuous and conditionally normally distributed, and a linear relationship exists between the observed and latent variables. Although this list of assumptions is not exhaustive, it does represent assumptions that are easily violated with itemlevel ordered-categorical data. Attempting to estimate model parameters, for example,
NIH-PA Author ManuscriptNIH-PA Author Manuscript NIH-PA Author Manuscript with dichotomous outcomes within the standard confirmatory factor model (as described by Jöreskog, 1969) results in parameter estimates that are biased and impossible to interpret accurately (Di-Stefano, 2002). However, all is not lost; just as logistic and ordinal regression techniques offer appropriate alternatives to linear regression when modeling dichotomous or polytomous (e.g., ordinal, Likert-type scales) outcomes, item factor analysis (IFA) offers an appropriate alternative to the common line...