Deciding on the Starting Number of Classes of a Latent Class Tree

Abstract

In recent studies, latent class tree (LCT) modeling has been proposed as a convenient alternative to standard latent class (LC) analysis. Instead of using an estimation method in which all classes are formed simultaneously given the specified number of classes, in LCT analysis a hierarchical structure of mutually linked classes is obtained by sequentially splitting classes into two subclasses. The resulting tree structure gives a clear insight into how the classes are formed and how solutions with different numbers of classes are substantively linked to one another. A limitation of the current LCT modeling approach is that it allows only for binary splits, which in certain situations may be too restrictive. Especially at the root node of the tree, where an initial set of classes is created based on the most dominant associations present in the data, it may make sense to use a model with more than two classes. In this article, we propose a modification of the LCT approach that allows for a nonbinary split at the root node, and we provide methods to determine the appropriate number of classes in this first split, based either on theoretical grounds or on a relative improvement of fit measure. This novel approach also can be seen as a hybrid of a standard LC model and a binary LCT model, in which an initial, oversimplified but interpretable model is refined using an LCT approach. Furthermore, we show how to apply an LCT model when a nonstandard LC model is required. These new approaches are illustrated using two empirical applications: one on social capital and the other on (post)materialism.

An empirically derived hierarchical tree typology of DSM-5 pathological personality traits in adolescence.

“…To ensure that potential classes would not be mere expressions of gender bias in the PID-5 items, we carried out a test of configural invariance using confirmatory factor analyses. Satisfactory configural invariance was found, indicating that the basic (van den Bergh et al, 2018) PID-5 item-to-facet structure was similar across gender groups (see for details).…”

“…Beginning at the root of the tree, LPA determines the number of classes at the first level of hierarchy (i.e., most general typology of individual profiles). This is assessed by estimating multiple standard LC models and subsequently examining the relative improvement in fit when adding more types (van den Bergh et al, 2018).…”

Key Issues and Potential Solutions for Understanding Healthcare Preference Heterogeneity Free from Patient-Level Scale Confounds

“…The standard information criteria tend to overstate the number, to include minor classes that are not policy relevant. A major new development in LCL is the ability to utilize a hierarchical tree to structure the latent classes with major policy-relevant classes being identified at the root of the tree and then split 5 further as needed to explain all the heterogeneity [15]. This innovative methodology was extended to SALC models by one of the presenters.…”

Classification of Hearing Aids Into Feature Profiles Using Hierarchical Latent Class Analysis Applied to a Large Dataset of Hearing Aids

“…Deciding on the number of classes that best fits the data is done in an exploratory setting. Depending on the dataset, it is not uncommon to end up with a large number of classes ( Van Den Bergh et al 2018 ). Such a result can become exceedingly complicated and difficult to interpret, reducing the practical use of the model.…”