Abstract
Latent class trajectory models (LCTMs) are often used to identify subgroups of patients that are clinically meaningful in terms of longitudinal exposure and outcome, e.g. drug response patterns. These models are increasingly applied in medicine and epidemiology. However, in many published studies, it is not clear whether the chosen models, where subgroups of patients are identified, represent real heterogeneity in the population, or whether any associations with clinically meaningful characteristics are accidental. In particular, we note an apparent over-reliance on lowest AIC or BIC values. While these are objective measures of goodness of fit, and can help identify the optimal number of subgroups, they are not sufficient on their own to fully evaluate a given trajectory model. Here we demonstrate how longitudinal latent class models can substantially change by making small modifications in model specification, and the impact of this on the relationship to clinical outcomes. We show that the predicted trajectory patterns and outcome probabilities differ when pre-specified cubic versus linear shapes are tested on the same data. However, both could be interpreted to be the " correct " model. We emphasise that LCTMs, like all unsupervised approaches, are hypotheses generating, and should not be directly implemented in clinical practice without significant testing and validation.