Determinants of the population health distribution, or why are risk factor-body mass index associations larger at the upper end of the BMI distribution?

Most epidemiological studies examine how risk factors relate to average difference in outcomes (linear regression) or odds a binary outcome (logistic regression); they do not explicitly examine whether risk factors are associated differentially across the distribution of the health outcome investigated. This paper documents a phenomenon found repeatedly in the minority of epidemiological studies which do this (via quantile regression) -associations between a range of established risk factors and body mass index (BMI) are progressively stronger in the upper ends of the BMI distribution. In this paper, we document this finding and provide illustrative evidence of it in a single dataset (the 1958 British birth cohort study). Associations of low childhood socioeconomic position, high maternal weight, low childhood general cognition and adult physical inactivity with higher BMI are larger at the upper end of the BMI distribution, on both absolute and relative scales. For example, effect estimates for socioeconomic position and childhood cognition were around three times larger at the 90 th compared with 10 th quantile, while effect estimates for physical inactivity were increasingly larger from the 50 th -90 th quantiles, yet null at lower quantiles. We provide potential explanations for these findings and discuss possible research and policy implications. We conclude by stating that tools such as quantile regression may be useful to better understand how risk factors relate to the distribution of health -particularly so in obesity research given conventional reliance on cut-points -yet for other outcomes in addition given the continuous nature of population health.