Dean Foster1,2, Howard Karloff, Kenneth E Shirley. 1. Department of Statistics, University of Pennsylvania, Philadelphia, Pennsylvania, USA. 2. Supply Chain Optimization Technologies, Amazon, Inc., New York City, New York, USA.
Abstract
OBJECTIVE: The objective was twofold: (1) to estimate for each individual the body mass index (BMI) which is associated with the lowest risk of death, and (2) to study variants of the BMI formula to determine which gives the best predictions of death. METHODS: Treating BMI = mass/height(2) as a continuous variable and estimating its interaction effects with several other variables, this study analyzed the NIH-AARP study data set of approximately 566,000 individuals and fit Cox proportional hazards models to the survival times. RESULTS: For each individual, a "personalized optimal BMI," the BMI for that individual which, according to the model, is associated with the lowest risk of death, is estimated. The average personalized optimal BMI is approximately 26, which is in the current "overweight" category. In fact, mass/height is a better predictor of death on the data set than BMI itself. CONCLUSIONS: The model suggests that an individual's "optimal" BMI depends on his or her features; "one-size-fits-all" recommendations may be not best.
OBJECTIVE: The objective was twofold: (1) to estimate for each individual the body mass index (BMI) which is associated with the lowest risk of death, and (2) to study variants of the BMI formula to determine which gives the best predictions of death. METHODS: Treating BMI = mass/height(2) as a continuous variable and estimating its interaction effects with several other variables, this study analyzed the NIH-AARP study data set of approximately 566,000 individuals and fit Cox proportional hazards models to the survival times. RESULTS: For each individual, a "personalized optimal BMI," the BMI for that individual which, according to the model, is associated with the lowest risk of death, is estimated. The average personalized optimal BMI is approximately 26, which is in the current "overweight" category. In fact, mass/height is a better predictor of death on the data set than BMI itself. CONCLUSIONS: The model suggests that an individual's "optimal" BMI depends on his or her features; "one-size-fits-all" recommendations may be not best.