St. James Ballroom
Determining Changepoints in Longitudinal Cardiovascular Health Scores Using Segmented Mixed Models (303867)
Norrina B. Allen, Feinberg School of Medicine, Northwestern University
Camilo F. Alonso, School of Public Health and Tropical Medicine, Tulane University
Lydia Bazzano, School of Public Health and Tropical Medicine, Tulane University
R. Sue Day, UT Health School of Public Health
Matthew Gillman, National Institutes of Health
Philip Greenland, Feinberg School of Medicine, Northwestern University
Markus Juonala, University of Tampere
Mika Kähönen, University of Tampere
*Amy Elizabeth Krefman, Feinberg School of Medicine, Northwestern University
Darwin Labarthe, Feinberg School of Medicine, Northwestern University
Tomi T. Laitinen, University of Turku
Terho Lehtimäki, University of Tampere
Lei Liu, Washington University of Medicine
Donald M. Lloyd-Jones, Feinberg School of Medicine, Northwestern University
Vito Michele Rosario Muggeo, Università degli Studi di Palermo, Italy
Katja Pahkala, University of Turku
Olli Raitakari, University of Turku
Linda Van Horn, Feinberg School of Medicine, Northwestern University
Larry S. Webber, School of Public Health and Tropical Medicine, Tulane University
Keywords: longitudinal, mixed models, changepoint, piecewise linear, cardiovascular, cohort
Ideal cardiovascular health (CVH) has been shown to decrease over time, but whether declines are constant over time is of particular interest for risk prevention. Piecewise linear mixed models with unknown changepoints can be challenging to fit without significant computational resources, or specification of starting values. CVH data were pooled together from five cohorts, including 18,290 individuals aged 8 to 55 years. An extension of the R package, ‘segmented’, was used to iteratively determine the fixed changepoint in a mixed model with a random participant intercept, adjusted for race, gender, and cohort. The data were subset by age to determine the existence of any remaining changepoints, and the best model was chosen using appropriate fit statistics. Two changepoints were found in the full adjusted model, at 16.8 years and 30.2 years. Estimates differ in stratified models by race (White: 16.2 years, Non-White: 18.3 years), and gender (Women: 25.3 years, 36.2 years; Men: 16.0 years, 41.2 years). This technique worked well for this large, longitudinal dataset where significant variation between individuals was present.