A comparison of unsupervised methods based on dichotomous data to identify clusters of airways symptoms: latent class analysis and partitioning around medoids methods.

Abstract

Latent class analysis (LCA) and partitioning around medoids (PAM) are popular data-driven methods for partitioning objects based on dichotomous data, remaining not clear which is better for large epidemiological datasets. Hence, we compared these methods in the identification of clusters of subjects with airways symptoms, using a large population-based data from the U.S. National Health and Nutrition Examination Surveys (NHANES). Adults from the NHANES 2007–2012 surveys were studied (n=18,619). We applied both LCA and PAM methods on self-reported lower airways (symptoms and related-outcomes), and hay fever. The two methods were compared according to: number of clusters, error rate and interpretability. The same number of clusters (five) was identified with both methods (Figure 1). Overall, it was obtained an error rate of 28% and all subjects included in the cluster without airways symptoms by PAM were also included in the same cluster identified by LCA. However, clinical interpretability was distinct between methods: PAM provided homogeneous classes and LCA identified divergent behavior. Although the two methods proved capable of recovering cluster structure based on dichotomous variables of a population-based study, they are not interchangeable. LCA appears to consider the heterogeneity of airways symptoms, however further testing is needed.