Modeling obesity histories in cohort analyses of health and mortality

Samuel H. Preston, Neil K. Mehta, Andrew Stokes

Research output: Contribution to journalReview articlepeer-review

Abstract

There is great interest in understanding the role of weight dynamics over the life cycle in predicting the incidence of disease and death. Beginning with a Medline search, we identify, classify, and evaluate the major approaches that have been used to study these dynamics. We identify four types of models: additive models, duration-of-obesity models, additive-weight-change models, and interactive models. We develop a framework that integrates the major approaches and shows that they are often nested in one another, a property that facilitates statistical comparisons. Our criteria for evaluating models are two-fold: the model's interpretability and its ability to account for observed variation in health outcomes. We apply two sets of nested models to data on adults age 50-74 years at baseline in two national probability samples drawn from National Health and Nutrition Examination Survey. One set of models treats obesity as a dichotomous variable and the other treats it as a continuous variable. In three of four applications, a fully interactive model does not add significant explanatory power to the simple additive model. In all four applications, little explanatory power is lost by simplifying the additive model to a duration model in which the coefficients of weight at different ages are set equal to one another. Other versions of a duration-of-obesity model also perform well, underscoring the importance of obesity at early adult ages for mortality at older ages.

Original languageEnglish (US)
Pages (from-to)158-166
Number of pages9
JournalEpidemiology
Volume24
Issue number1
DOIs
StatePublished - Jan 2013
Externally publishedYes

ASJC Scopus subject areas

  • Epidemiology

Fingerprint Dive into the research topics of 'Modeling obesity histories in cohort analyses of health and mortality'. Together they form a unique fingerprint.

Cite this