Page 105 - Contributed Paper Session (CPS) - Volume 3

CPS1954 Vincent C. et al.
faltering and then the HAZ remains relatively stable after that. Generally,
subgroup 2 has a longer duration of faltered growth. Children in subgroups 3
and 7 have a steep decline in the HAZ score before a short interval of recovery
is observed, and then followed by another onset of faltered growth. The
differences between these two subgroups lie in the time point at which growth
catch up takes place and the rate of improvement. The growth catch up phase
for children in subgroup 3 is milder and happens between t = 0.25 and
t = 0.50 whereas huge jumps in the HAZ happen between t = 0.50 and t = 0.80
in subgroup 7. Another pair of clusters which are similar is subgroups 4 and 8
whereby the growth looks like a sinusoidal wave, but with different amplitudes.
The HAZ for children in subgroup 5 reaches a peak and then starts to decline.

4. Conclusion
In this article, we use the broken stick model as the basis to propose an
approach which incorporates a classifier within a regression model. This allows
the classification of growth curves into different patterns based on the vectors
of regression parameters to be achieved within a single model. In order to
capture the heterogeneity in the growth velocity between children, we extend
the broken stick model to allow for mixture distributed random slopes. The
classification of an individual child’s growth profile is then determined by the
component of mixture distribution from which the vector of velocities derived
from the regression model is generated. We model the distribution of growth
velocity non-parametrically in a Bayesian framework using a DP prior. The DP
prior adapts complexity of the model to the amount of data available without
having to choose the number of mixture components, which is often unknown
in practical applications. Another contribution of this paper is to introduce the
idea of random change points into the broken stick model in order for the
knots to adjust their locations instead of having them fixed. These change
points are modelled as random effects so that the difference in the timing of
growth phase between children is taken into consideration in the model.

94 | I S I W S C 2 0 1 9