Abstract
The EU-SILC database contains annually gathered rotating-panel data on a household level covering indicators of monetary poverty, severe material deprivation or low work household intensity. Data are obtained via questionnaires leading to outcome variables of diverse nature: numeric, binary, ordinal or general categorical.
In our previous contribution to MME 2020 we presented a clustering method for such a type of data. The used thresholding approach of latent numeric counterparts of binary and ordinal outcomes suffered from slow convergence and unclear interpretation of resulting estimates.
Hence we propose an alternative approach which again exploits a Bayesian variant of the model based clustering (MBC). Nevertheless, the underlying models are all of a generalized linear mixed model (GLMM) nature: (proportional odds) logit model for (ordinal) or binary indicators, multinomial logit model for general categorical outcomes and a standard linear mixed model for numeric outcome.
Czech households interviewed within the EU-SILC project between 2005 and 2018 are then divided into several groups of similar evolution of income, housing costs, self-evaluations and other indicators.