Where Is the Geography? A Study of the Predictors of Obesity Using UK Biobank Data and Machine Learning

Abstract

AbstractIn this study, we adopted individual baseline data from the UK Biobank cohort of participants aged 40–69 across the UK to explore whether there is evidence of the geography related to health disparities in obesity. First, we used multilevel models to decompose the variation in body mass index (BMI) values to examine the presence of spatial clustering patterns of individual BMI values at various geographic scales. Next, we looked at whether key predictors of obesity, such as physical activities and dietary habits, differ across 6 cities in England by using a machine learning approach. To do this, we trained random forest models in one city, and we used them to predict BMI values in other cities to see if the models were spatially transferable. Subsequently, we turned to explore socio-economic status, which is one of the direct interests in the literature with obesity and used those in combination with multilevel models to check for the existence of spatially varying effects. The results of the multilevel null models indicate that most of the variance of BMI is due to individual variation, suggesting little evidence of geographical clustering at any geographical scales. The machine learning prediction results show that the effects of the main identified risk factors for obesity are stable (spatially stationary) across cities, based on approximately the same predictive power and broadly constant effect sizes of main factors. Multilevel models taking socio-economic status into account further support that individual and neighbourhood deprivation levels display limited geographical variation in their effects on obesity across the study areas. Contrary to our expectations, the models together suggest the limited association of geographical context with obesity, among the UK Biobank participants.