Assessing the Role of Latent Variables in Modelling and Predicting the Determinants of Obesity

A Structural Equation Modelling Approach

Résumé

Obesity is a complex and multifactorial public health problem driven by the interaction of biological, behavioural, and environmental determinants. Accurately modelling these interdependencies remains methodologically challenging, particularly when conventional regression and machine-learning approaches prioritise prediction at the expense of interpretability. This study evaluates the contribution of latent variables to obesity modelling and prediction using a comparative structural equation modelling (SEM) framework. Using survey data from 2111 individuals aged 14 to 61 years, two SEM specifications were estimated: (i) a construct-based model incorporating three latent determinants namely, Demographic and Anthropometric Factors (DAF), Dietary and Eating Behaviour (DEB), and Lifestyle and Physical Activity (LPA), together with family history of overweight, and (ii) an observed-variable model in which all predictors were entered directly.
Results indicate that both models capture meaningful relationships with obesity level (NObeyesdad), but they differ substantially in explanatory power and interpretability. The construct-based model explains 70.9% of the variance in obesity and identifies DAF, DEB, and family history as the dominant drivers, with lifestyle factors exerting weaker effects. In contrast, the observed-variable model explains 95.9% of the variation in obesity, largely driven by direct anthropometric indicators, especially weight, and multiple behavioural variables. However, this model is considerably more complex and mixes determinants with outcomes and correlates, raising concerns regarding overfitting and theoretical coherence.
Overall, the findings demonstrate a clear trade-off between predictive accuracy and conceptual clarity. Latent variable modelling provides a parsimonious and theory-consistent representation of obesity determinants, while observed-variable SEM maximises statistical fit. These results suggest that for predictive purposes, the model without constructs is better.