Comparison of two bootstrap procedures in the case of hidden Markovian model clustering

TitreComparison of two bootstrap procedures in the case of hidden Markovian model clustering
Type de publicationConference Paper
Year of Publication2016
AuteursTaushanov, Z, Berchtold, A
PublisherEuropean Regional Section of the IASC
Place PublishedOviedo, Spain
Mots-clésbootstrap, clustering, hidden Markov model, mixture models

The clustering of longitudinal data sequences is considered using a latent Markovian model (HMTD) combining Gaussian distributions and covariates. The main objective is to evaluate the significance of the estimated parameters. At first, different model specifications are optimized and the one providing the best clustering in terms of BIC is selected. Two different bootstrap procedures are then applied and compared in order to investigate the significance of the parameters of this optimal solution. First, a standard bootstrap procedure is applied using the full original sample and the optimal model with multiple components (clusters) is computed at each iteration. That leads to solutions with different degrees of similarity with the optimal solution and the well-known label-switching problem may occur. An alternative procedure is proposed that consists in applying separate bootstrap procedures on each subsample defined by the optimal clustering. In this case, a single component model is estimated from each bootstrap iteration and for each cluster separately. This method also provides a confidence interval for each parameter and avoids the label-switching problem. The pros and cons of each approach are described and examples based on real data are provided.

Citation Key2420
Refereed DesignationUnknown