An efficient Fisher-scoring algorithm for fitting latent class models with individual covariates

摘要

For latent class models where the class weights depend on individualcovariates, we derive a simple expression for computing the score vector and aconvenient hybrid between the observed and the expected information matriceswhich is always positive defnite. These ingredients, combined with amaximization algorithm based on line search, provides an efficient tool formaximum likelihood estimation. In particular, the proposed algorithm is suchthat the log-likelihood never decreases from one step to the next and thechoice of starting values is not crucial for reaching a local maximum. We showhow the same algorithm may be used for numerical investigation of the effect ofmodel mispecifications. An application to education transmission is used as anillustration.