Essays on Selection in Health Behaviors

Abstract

Lifestyle factors are an important health input, but estimating causal effects from observational data is challenging because take-up of health behaviors is non-random. This dissertation examines how this problem---sometimes referred to as selection or confounding---affects the reliability of observational estimates in nutritional epidemiology. In the first chapter, I show that changes in supplement take-up following new health research are accompanied by changes in selection on both observable and unobservable dimensions. Even after controlling for observable confounders, estimated associations between supplement use and health outcomes strengthen substantially during high-take-up periods. I provide evidence that this effect is driven by adopters: individuals who begin supplement use in response to changing guidance. In the second chapter, I leverage information on changes in observable confounders and individual behavior over time to control for residual selection bias. I find that selection on observables is not a reliable predictor of selection on unobservables, complicating bias correction and bounding exercises. Directly controlling for adopter status is more effective, highlighting the importance of identifying the source of variation being used for estimation. In the third chapter, I document that measurement error in self-reported food intake and physical activity data varies systematically across groups: the discrepancy between self-reported and monitor-recorded physical activity is larger for smokers and higher-BMI individuals. Calculated energy balance from self-reported data is nearly uncorrelated with implied energy balance, and exhibits a negative relationship: on average, individuals who report larger caloric deficits gain more weight over the next two years. I describe how this type of non-random measurement error can affect observational studies, not only on the effect of diet and exercise, but also other studies controlling for these variables as confounders. Taken together, these results demonstrate that selection bias in observational health research remains a serious concern, even in well-designed prospective cohort studies.