21st Jun 2019
Introduction: Life-course approaches to determine risk factor exposure are increasingly favoured over traditional ‘once-only’ epidemiological determination. One is latent class trajectory modelling (LCTM), which clusters individuals’ changes in exposure over time and can be used as a tool for identifying early divergent adverse trajectories. There is increasing use of LCTMs in mainstream epidemiology, but often with poor model description, and an over-reliance on BIC for model selection. Here, we aimed to explore methods in deriving and validating LCTMs from multiple cohorts to ensure that they properly represent observed patterns across different populations.
Methods: We interrogated three cohorts, AARP (N: 321,827), PLCO (N: 147,488) and WHI (N: 151,363), with longitudinal BMI as the exposure and cancer incidence as the endpoint of interest. We extended our previous work (https://bmjopen.bmj.com/content/8/7/e020683) to (i) testing multiple start points to obtain the global maximums; (ii) facilitating model choice with alternative metrics to BIC; (iii) developing visual model assessment tools.
Results: We illustrate a number of examples where deviation from model assumptions yield very different classifications. After arriving at preferred models, we show that LCTM improve the performance characteristics of BMI exposure, compared with once only BMI measures, however, this improvement is clinically modest. LCTM might best identify specific sub-populations that have very high risk for cancer incidence.
Conclusions: The study highlights that model selection needs to be undertaken with care and not based solely determined by lowest BIC. We emphasise that multiple start points should be tested when using these models.