The sensitivity of latent class analysis
Description
Latent class analysis was originally introduced by Lazarsfeld to analyze categorical latent variables from a set of categorical manifest variables. Several test statistics have been previously proposed for evaluating the goodness of fit of latent class model. They provide the means to test whether a latent factor explains the observed associations. So far, however, little is known about the sensitivity of latent class analysis to detect the true latent structure (convergent) or to discriminate an inappropriate model from the true model (discriminant). This study provides some strategies for researchers to assess both convergent and discriminant sensitivity of latent class analysis using simulation methods The impact of several factors on convergent sensitivity of latent class analysis were investigated. These factors include (1) sample size, (2) nearness of unconditional probabilities to 0.5, (3) nearness of conditional probabilities to one or zero, (4) strength of association between latent and manifest variables and (5) nearness of the pair of conditional probabilities among all latent classes for any manifest variable. Factors (1), (2) and (3) address the sample size issue. Factors (4) and (5) address the strength of association. Results confirm that convergent sensitivity is globally improved by increasing either the sample size or the strength of association It was also found that the discriminatory power of latent class analysis to distinguish an inappropriate model from the true latent structure is significantly high. Based on the findings of this study, we believe researchers can significantly improve their study designs