Abstract:
|
Classification of longitudinal data and prediction of future outcomes is a growing area of research interest. Recently introduced random effects expectation-maximization (RE-EM) tree accounts for within-subject correlational structure in longitudinal data and fits different tree structures for different periods when splits on time. The method can be applied to unbalanced panels and has shown better performance in predicting data in future periods. In this study, we applied this method as a data mining tool to a dataset of early childhood growth patterns of 488 children identified as obese at the age of 5 years. Application of this method with random intercept and slope divides the age of the children into three splits (3.05, 30.95, and 41.25 months) to detect homogeneous pieces of trends of growth patterns. We used splits as knots to fit a piecewise linear mixed effects model and applied the posthoc-mixture model of BLUPs of random coefficients to classify children in homogeneous groups (2-5) with respect to their pathways to obesity. Based on several evaluations, we identified 5 group-classification as optimum to represent the heterogeneous pathways of obesity in this dataset.
|