A Dynamic Learning Model of Smoking Behavior

Presenter: Michael Darden, University of North Carolina, Chapel HIll

Abstract

Do individuals learn from the "warning signs" of several major chronic conditions commonly associated with smoking? This study examines how smokers use health information to update their expectations prior to major health shocks. Indeed, most papers that have studied learning or expectation formation with respect to smoking and health have examined changes after the occurrence of major health shocks. Furthermore, most studies restrict their focus to those individuals above the age of fifty. Using rich panel health and smoking data from the Framingham Heart Survey, this paper estimates the structural parameters of a dynamic discrete choice model that incorporates learning. The structural approach, in addition to the long panel dataset, allow for the examination of behavior across the life cycle. The key learning aspect of the model is the observed heterogeneity in health marker responses to smoking. An individual has some belief as to their idiosyncratic effect of smoking on a set of health markers. This belief is updated each period with information from a health exam. I place a Bayesian structure on the learning process and treat an individual's belief as a state variable in the dynamic optimization problem. The forward-looking agent uses the health exam information, in conjunction with prior beliefs, to forecast future health outcomes conditional on current smoking decisions. Individuals choose a sequence of smoking decisions to maximize the expected present discounted value of lifetime utility. The relevant tradeoff is current utility from smoking versus potentially negative future health consequences.

This paper also addresses the issue of "smoking depreciation". To what extent does the health of a smoker converge to that of a nonsmoker given smoking cessation? To answer this question, this paper treats cigarette smoking as a capital stock. I construct the stock of smoking to be a weighted average of several smoking variables using Principal Component Analysis. Examples of variables included are duration of smoking, intensity with which one smoked, and years of smoking cessation. Each period, the stock variable updates along with other state variables of the model. The stock evolves as a function of it's lagged value and the lagged smoking decision.

The World Health Organization estimates that smoking related-diseases kill one in ten adults globally. In the United States, illness from smoking is estimated to add $157 billion per year to national health expenditures. The clear policy goal is to promote smoking cessation. Given the attention paid to anti-smoking advertising campaigns, this study sheds light on the importance of personalized health information. Preliminary results suggest that smokers do respond changes in acute measures of health. Personalized health information appears to be a powerful tool with which to encourage smoking cessation.