Title: Bayesian multilevel hidden Markov models: A reliability guideline for applied researchers
Authors: Sebastian Mildiner Moraga - Utrecht University (Netherlands) [presenting]
Emmeke Aarts - Utrecht University (Netherlands)
Abstract: The popularity of hidden Markov models (HMMs) in the econometric, social and behavioural literature has been steadily increasing lately. Moving beyond standard statistical tests, HMMs offer a statistical environment to optimally exploit the information present in real-time behavioural data, for example, uncovering the dynamics of behaviour. For this type of data, multilevel hidden Markov models (MHMMs) are a particularly good fit: they allow for the accommodation of variability between subjects and the estimation of both subject-level and group-level parameters. To create guidelines for applied researchers, we assessed the effect of three researcher controlled factors -the number of subjects, the number of occasions and the number of dependent variables used to train the model- and two data quality conditions -noisiness and overlapping of the conditional distributions- on the estimation performance of a multinomial Bayesian MHMM. Our results reveal that increments in the number of subjects and the number of dependent variables included are more beneficial for the estimation performance than increments on the number of occasions. These effects are consistent across the levels of complexity in the conditional distributions. We conclude that measuring multivariate rather than univariate data results in the most cost-effective gains in estimation performance and the likelihood of convergence.