Title: Evaluation of the robustness of stepwise latent class estimators and a new two-stage estimator
Authors: Zsuzsa Bakk - Leiden university (Netherlands) [presenting]
Abstract: The aim focuses on classification error corrected stepwise estimation approaches of Latent Class (LC) models with external variables. Currently two approaches are available, ML and BCH, that follow a similar procedure: in the first step the underlying latent construct is estimated based on a set of observed indicator variables, next, in step two, cases are assigned to the LCs, and, finally, in the third step, the class assignments are used in further analyses, while correcting for classification error. We discuss the robustness of the stepwise estimation procedures to different violations of underlying model assumptions and highlight that when the presence of direct effects between the external variable and the indicators of LC model are ignored, both approaches lead to biased estimates in the last step. We propose an alternative two-stage estimator to address this problem. Using this two-stage estimator in step one, a LC model is estimated with the indicators only, and in the second step, the external variable of interest is added to the model, freely estimating the association between the external variable and LC membership, while keeping the parameters of the measurement model fixed to the values estimated in step one. Using this approach, local fit measures can be used to test which fixed effects need to be freed to account for the presence of direct effects.