The impact of ordinal scales on Gaussian mixture recovery

Gaussian mixture models (GMMs) are a popular and versatile tool for exploring heterogeneity in multivariate continuous data. Arguably the most popular way to estimate GMMs is via the expectation-maximization (EM) algorithm combined with model selection using the Bayesian information criterion (BIC). If the GMM is correctly specified, this estimation procedure has been demonstrated to have high recovery performance. However, in many situations, the data are not continuous but ordinal, for example when assessing symptom severity in medical data or modeling the responses in a survey. For such situations, it is unknown how well the EM algorithm and the BIC perform in GMM recovery. In the present paper, we investigate this question by simulating data from various GMMs, thresholding them in ordinal categories and evaluating recovery performance. We show that the number of components can be estimated reliably if the number of ordinal categories and the number of variables is high enough. However, the estimates of the parameters of the component models are biased independent of sample size. Finally, we discuss alternative modeling approaches which might be adopted for the situations in which estimating a GMM is not acceptable.

Automated summary

Gaussian mixture models (GMMs) are widely used for exploring heterogeneity in multivariate continuous data, but their performance in estimating GMMs for ordinal data is uncertain. In this study, we investigate this by simulating data from various GMMs, thresholding them in ordinal categories, and evaluating recovery performance. We find that the number of components can be reliably estimated with enough ordinal categories and variables, but the estimates of component model parameters are biased regardless of sample size.