# Modelling and Predicting Population‐Level Growth With Individual‐Level Information

## TL;DR

This study uses individual growth data to improve predictions of children's height and weight at the population level.

## Contribution

The study introduces three approaches for joint prediction of height and weight using Bayesian hierarchical modeling with individual-level data.

## Key findings

- Using individual-level data improves prediction accuracy for weight and BMI in children.
- Bayesian hierarchical models reduce divergence from observed measurements in growth predictions.
- Multiple prediction checks are essential to evaluate the strengths and flaws of each approach.

## Abstract

The development of height, weight, and body mass index (BMI) in children has been the subject of considerable interest due to secular changes in growth patterns, such as increases in height and rising obesity rates. Predicting growth in a target population is particularly challenging when the population comprises of individuals with and without past growth data. In this study, we present three approaches for the joint prediction of height and weight in that situation. The predictive performance of each approach is evaluated using a range of measures that assess different properties of the prediction distributions. We also compare the approaches to interpret their clinical relevance, particularly in terms of prediction accuracy. The developed prediction approaches vary in their use of past growth data. We predict growth for a target population of children aged 4–11 years in 2021, residing in three municipalities in Finland. We employ longitudinal register data on height and weight, collected from children aged 2–11 years between 2014 and 2020 in these municipalities to construct a Bayesian hierarchical linear model (HLM) for growth prediction. Additionally, we estimate posterior unconditional distributions of height, weight, and BMI for within‐sample model validation. The inclusion of individual‐level data in the predictions reduced the divergence from observed measurements, particularly for weight and BMI. This is important given the skewed distribution of the measurements with increasing age. Incorporating individual‐level information is also beneficial for child‐specific predictions. Our study highlights the importance of multiple prediction checks to understand the flaws and strengths of each prediction approach.