Clustering Ordinal Data via Latent Variable Models
Item response modelling is a well established method for analysing ordinal response data. Ordinal data are typically collected as responses to a number of questions or items. The observed data can be viewed as discrete versions of an underlying latent Gaussian variable. Item response models assume that this latent variable (and therefore the observed ordinal response) is a function of both respondent specific and item specific parameters. However, item response models assume a homogeneous population in that the item specific parameters are assumed to be the same for all respondents. Often a population is heterogeneous and clusters of respondents exist; members of different clusters may view the items differently. A mixture of item response models is developed to provide clustering capabilities in the context of ordinal response data. The model is estimated within the Bayesian paradigm and is illustrated through an application to an ordinal response data set resulting from a clinical trial involving self-assessment of arthritis.
KeywordsLatent Trait Ordinal Data Item Parameter Marginal Likelihood Markov Chain Monte Carlo Algorithm
This work has emanated from research conducted with the financial support of Science Foundation Ireland under Grant Number 09/RFP/MTH2367.
- Frühwirth-Schnatter, S. (2004). Estimating marginal likelihoods for mixture and Markov switching models using bridge sampling techniques. Statistica Sinica, 6, 831–860.Google Scholar
- Meng, X. L., & Wong, W. H. (1996). Simulating ratios of normalizing constants via a simple identity: a theoretical exploration. The Econometrics Journal, 7, 143–167.Google Scholar