Tilburg School of Social and Behavioral Sciences
Prof. J.K. Vermunt & Dr J. Mulder
On November 13th 2017, Geert van Kollenburg will defend his thesis entitled
Computer Intensive Methods for Evaluating Latent Class Model Fit
(Summary not available)
Diagnostics for latent class models
I am working on a project funded by the NWO Vici grant “Stepwise model-fitting approaches for latent class analysis and related methods”.
Assessment of model fit traditionally involves calculating p-values based on asymptotic reference distributions. However, this is not always appropriate or possible. When contingency tables are sparse, asymptotic reference distributions may lead to dramatically biased Type-I-error rates (Reiser & Lin, 1999).
In other situations, statistics are used for which the sampling distribution is unknown. In these situations empirically derived sampling distributions can be obtained through resampling techniques.
In my first paper we applied the posterior predictive check (Gelman et al., 1996; Rubin & Stern, 1994) to obtain empirical p-values for a number of commonly used fit statistics within the context of latent class analysis.
In my second paper we are developing a calibration method for the posterior predictive check. The rest of my project will focus on developing diagnostics for latent class models in combination with resampling techniques.
NWO project at Tilburg University