Latent class growth analysis for ordinal response data in the Distress Assessment and Response Tool: an evaluation of state-of-the-art implementations

TL;DR

This study compares different software implementations of latent class growth analysis for ordinal data, highlighting their strengths and limitations using simulated and real patient data from a cancer center.

Contribution

It provides a systematic evaluation of LCGA implementations, offering practical recommendations for analyzing ordinal response data in health research.

Findings

Mplus and lcmm achieve high classification accuracy

Proc Traj overestimates classes and fails to converge

Mplus is faster but limited to 10 levels

Abstract

Latent class growth analysis is a popular approach to identify underlying subpopulations. Several implementations, such as LCGA (Mplus), Proc Traj (SAS) and lcmm (R) are specially designed for this purpose. Motivated by data collection of psychological instruments over time in a large North American cancer centre, we compare these implementations using various simulated Edmonton Symptom Assessment System revised (ESAS-r) scores, an ordinal outcome from 0 to 10, as well as the real data consisting of more than 20,000 patients. We found that Mplus and lcmm lead to high correct classification rate, but Proc Traj over estimated the number of classes and failed to converge. While Mplus is computationally faster than lcmm, it does not allow more than 10 levels. We therefore suggest first analyzing data on the ordinal scale using lcmm. If computational time becomes an issue, then one can group the scores into categories and implement them in Mplus.