Weighting the Factors Associated with Children Obesity: A Random Forest Approach

Obesity and overweight (OWO) are a recognized worldwide health issues. Individual data are needed in order to assess role played by different factors. A key point is whether commonly accepted risks factors for child obesity play the same role in the various countries. Difficulties arise when the number of variables/number of subjects ratio is close to, if not greater than, one. This makes common regression approaches impracticable.

Data on 2640 children 6-11 years (Argentina, Brazil, Chile, France, Georgia, Germany, UK, India, Italy and Mexico) were collected on more than 90 parameters (anthropometrics, built environment, familiar socio economic status, food and activity frequency). Two outcomes were considered, BMI z-scores and WHO classes of OWO vs. Normal children.

Given that sample size is heterogeneous across countries (from 60 up to 1640 children), role played by each potential factor associated with both outcomes separately was estimated using Random Forests (RF), which have been implemented using 150000 bootstrap samples using Bylander’s bias-correction approach. All factors have been used as potential predictors of both outcomes. One-hundred permutations per tree were run for assessing each factor’s importance, using the mean of squared residuals for BMI z-score and the Out-of-Basket (OOB) classification error rate for OWO vs. Normal. Factors do not explain variability in BMI z-score (from 5% up to 34%) and to classify OWO children (error rate from 5.50% up to 94.7%), in the same extent in the various countries, suggesting that cultural heterogeneity exists in the roles played by the same factors on children’s obesity.