Numerical Examination on Sensitivity in Latent Class Analysisby Andersen's Diagnostics －31－ Numerical Examination on Sensitivity in Latent Class Analysis by Andersen's Diagnostics

Under the assumption of the local independence, latent class analysis can be reduced to a parametric multinomial distribution. For sensitivity analysis in a multinomial dis-tribution model, in general, a needed diagnostics is the diagnostics that measures the influence by deleting a single cell, rather than an observation. In this paper we apply Andersen's diagnostics, which is suggested for a paramet-ric multinomial distribution, to two artificial latent class models and data set analysed by Stouffer and Toby, and examine their numerical results. Also it is shown that the reliability of estimated parameters in latent class analysis is evaluated by thier asymptotic variances at ML estimates.