This repository contains materials for the development of a novel method for modeling infant growth trajectories - a primary goal of infant development researchers. This model seeks to facilitate discovery of latent classes of infant trajectories using finite mixtures of multivariate skew normal distributions. We allow for covariates on the probabilities of latent class membership using Bayesian multinomial regression. The multivariate skew normal distribution is chosen to account for positive correlation in longitudinal measurements of infant development, which are often skewed due to their standardization to a global reference population. This method also includes to ability to impute missing outcomes using draws from a conditional multivariate skew normal distribution, which we show to be a superior imputation scheme as compared to multiple imputation or complete case analysis. Our method makes use of data augmentation to allow for all parameters to be updated with easily sampled Gibbs steps, making our MCMC scheme highly accessible. We apply our method to data from the Nurture study - a recent cohort of infant-mother pairs for whom development data was collected throughout the first year of life.

Manuscript

Supplement

Poster

Simulation

Application