Comparison of growth curve models for assessing height in a South African birth cohort

Abstract

Childhood malnutrition is a major concern in low- to middle- income populations. This dissertation uses longitudinal data on height measurements of babies between 0 and 4 years of age to construct growth curves, which serve as a tool for assessing the health and nutritional progress of children. We wish to characterise the way height changes over time and identify predictors of that change. Various mixed effect models were fit and compared to neural networks in terms of model fit, interpretability of parameters as well as predictive power. The best fitting mixed-effect model was the Berkey-Reed 2nd order model. The neural network compared well with this model, indicating that neural networks may serve as a useful alternative to modelling longitudinal growth data. Subsequently, logistic regression was used to explain the relationship between various pre- and post-natal risk factors for stunting, a shortfall in height relative to age. The results were compared to a random forest model. Methods for variable importance in classification problems using tree-based methods were explored. The random forest model appeared to perform similarly to the logistic regression model in terms of predictive power and variable interpretation. This dissertation contributes in investigating the possibility of using machine learning techniques to identify probable correlates of childhood malnutrition.