Applied Computational Modeling Approaches in Cigarette Smoking Epidemiology: Expanding Statistical Associations to Convey Theoretical Pathways

Abstract

50 years since the landmark 1964 Surgeon General’s report on smoking and health, cigarette smoking remains the leading preventable cause of death and disability in the United States. The success of epidemiology and public health in the study of cigarette smoking, both as an exposure as well as a health outcome, has offered rich datasets and mechanistic discoveries that provide opportunities for the evolution of epidemiologic methods. Specifically, advancing computational science approaches allow for novel applications of methodologies, such as agent-based modeling or networks theory, in the epidemiological sciences to expand on existing knowledge. In this dissertation, we utilize approaches from epidemiology, statistics, computer science, and the philosophy of science to explore a range of hypothesized dynamics of smoking behavior that could contribute to changes in population-level smoking prevalence.

We begin with a computational model that weighs the magnitude of the potential harms and benefits of electronic cigarette (e-cigarette or vaping) use from an adult smoking prevalence standpoint. We find that e-cigarettes can exert a much larger influence on smoking prevalence through routes of smoking cessation, as opposed to smoking initiation, if e-cigarette use remains primarily concentrated among current smokers. Conversely, e-cigarettes would need to behave as extremely effective gateways for smoking initiation, and never smokers would need to become e-cigarette users at substantially higher levels than currently observed, for these products to independently generate increases in population-level smoking prevalence.

Next, we explore how contextual and individual network factors and demographic covariates change the effect of peer influence on smoking behavior in the National Longitudinal Study of Adolescent to Adult Health (Add Health). Using stratified mixed effects models, we find that the magnitude of friendship influence on smoking initiation differs by school social network density. We additionally find that the contextual factors, rather than peer influence, may be stronger predictors of smoking cessation. The effect estimates of these factors on smoking cessation of also differ by network density.

Extending these results, we conclude with an abstract simulation of the hypothesized mechanisms that contribute to the outcomes of the stratified mixed effects model described previously. We find that network structure and peer influence are sufficient in combination to generate substantial differences in smoking prevalence by urbanicity, sex, and race, among US adolescents. These results provide evidence that support the potential for effect modification by network density on the hypothesized pathway between friendship influence and smoking behavior.

While the field of tobacco control has been traditionally amenable to computational modeling approaches, few studies use computational modeling within an epidemiologic framework to provide support for hypothesized causal pathways that contribute to smoking behavior outcomes. Such perspectives are critical as the tobacco landscape continues to change with the introduction of new products, and as we gain a better understanding of the role that social networks play in the propagation of health behaviors. Through the integration of statistics, computational modeling, and epidemiologic methods, this dissertation seeks to provide insights into the potential causal pathways between various risk factors and smoking behavior outcomes. The results and discussions of this dissertation present potential avenues through which computational modeling can contribute added value to epidemiologic methods, in addition to our understanding of smoking behavior, beyond those of projection and evaluation.