The problem of estimation and forecasting of obesity prevalence using sparsely collected data

Abstract

The problem of estimation and forecasting of population nutritional status has been addressed in the literature, showing successful results when the data are available and frequently collected over time. However, most low and middle -income countries collect nutritional status data sparsely, and consequently, the uncertainty/absence of information may negatively affect decisions from policymakers. In this context, the problem of estimation and forecasting of obesity prevalence using sparsely collected cross-sectional data is formally stated and a novel sequential approach to address it is proposed. Specifically, this work describes the nutritional status dynamics using a system of nonlinear difference equations, where the set of transition probabilities are unknown parameters due to the sparsely collected cross-sectional data. Then, an artificial neural network alike model is proposed through its equivalent nonlinear programming model, considering the difference equations system as constraints as well as bounds for the transition probabilities based on literature data. In addition, comprehensive data collection and information analysis processes to compute demographic parameters are defined. As the model is non -convex, an optimal solution is characterized and coined as stable; and thereafter assessed in terms of its goodness -of -fit. Computational experiments and a resolution scheme using a rollinghorizon forecasting/back-casting approach and divergence metrics is proposed. To illustrate the usefulness of this novel approach, Chile is used as a case study. Results show an accuracy up to 90%, forecasting the men and women obese population (BMI >= 30.0 kg/m2) for 2024, reaching 30.6% (95% CI: 28.4-32.8%) and 32.6% (95% CI: 29.1-36.0%), respectively.