From the University of Toronto, Toronto, Ontario, Canada.
Correspondence: Gail McKeown-Eyssen, Department of Public Health Sciences, Faculty of Medicine, University of Toronto, 6th Floor, 155 College Street, Toronto, Ontario, Canada M5M 3T7. E-mail: email@example.com.
John Potter1 has asked epidemiologists to think broadly about factors responsible for obesity. Although energy balance is important, he recognizes the need for research also to consider the role of genes and of such diverse social factors as advertising, transportation, and eating disorders. To implement Potter's suggestions, it is necessary to design studies to consider how factors in the social and physical environments combine with personal behavior and genetics to determine obesity and its health consequences. Such studies need statistical methods that permit pathways between relevant factors to be examined. Structural equation modeling is a method that permits such analysis.2 The technique, although well known to social scientists, has not been widely adopted in epidemiology.
MODELING PATHS TO OBESITY
With structural equation modeling, the investigator can specify sequential relationships among variables. For example, it might be proposed that a social factor such as income influences diet and physical activity and that each of these in turn influences obesity (Fig. 1). Diet and physical activity are therefore proposed as “mediators” of the effect of income on obesity. In addition, there might be a direct effect of income on obesity without specified intermediate factors. These 5 relationships are shown by arrows in Figure 1. The aspects of diet considered in this example are intake of fruit and vegetables combined into a single so-called latent variable using factor analysis. The structural equations model estimates all relationships simultaneously. In the example, structural equation modeling used with data from the Ontario Health Survey models the relationships of income to diet and to physical activity with linear regression and models the relationships of diet, physical activity, and income to obesity with logistic regression—comparing the obese to those of normal weight.3 The overall model fits the observed data as indicated by the small χ2 statistic. T-statistics corresponding to the regression coefficients from the models are shown in Figure 1. In these data, income is positively related to diet and diet is negatively related to obesity; thus, diet mediates the effect of income on obesity. In addition, there is a direct negative relationship between income and obesity with no intermediate variables identified in the model. Physical activity is not a mediator of the relationship between income and obesity, both because the association of income to physical activity is weak, and because there is no substantial relationship between physical activity and obesity. Thus, structural equation models allow examination of the relationships among all variables and do not dismiss some associations as confounding. Understanding the complex interplay of factors involved in the etiology of obesity will benefit from such analyses.
HOW DO GENES FIT IN?
Potter1 suggests that it is important to consider the role of genetics in determining adiposity. Careful consideration of the nature of analytic models is needed in such studies. Genes are often considered as effect modifiers of environmental factors, and such gene–environment interactions are assessed through multiplicative terms in regression models. However, there may be an additional way in which genes act. Is it possible that an environmental factor such as diet might influence gene expression and that the resultant cascade of biologic events could lead to the development of obesity? In such a situation, genes and their biologic consequences might be seen as mediators of the effect of the dietary environment, not as effect modifiers. Mediation cannot be modeled by multiplicative interaction terms, but structural equation models can assess such relationships.
Thus, Potter's1 charge to researchers to look broadly at the full range of social, personal, and genetic factors that may influence obesity requires appropriate methods of modeling relationships. To do this, it may be necessary to move beyond familiar techniques to methods such as structural equation modeling that may provide a more comprehensive picture of the complex pathways to obesity.
FROM OBESITY TO DISEASE
In addition to considering paths to obesity, it will also be important to develop a sound understanding of the paths from obesity to disease. Obesity is associated with the risk of many chronic diseases, including hypertension, heart disease, diabetes, and cancer at a number of sites. Knowledge of these associations has been developed by studies that have focused, by design or during analysis, on single disease end points. Recognition of similarities among the associations of various diseases with diet, physical activity, and obesity has led to suggestions of biologic pathways that might account for associations of the same risk factor with a variety of health outcomes.
However, one might still wonder what determines the biologic path, and hence the diseases, to which exposed individuals succumb. Could it be that lifestyle and subsequent obesity set up the body's biologic systems for failure, and that the particular biologic system that fails is heavily influenced by the individual's genetic milieu (Fig. 2)? What study designs and analysis would be suitable to explore such a hypothesis? Could a metaanalysis of case–control or cohort studies of different diseases be used? Or would it require a cohort study, adequately powered for a number of separate health outcomes, to be analyzed simultaneously for all relevant diseases?
In the 19th century—a mere moment ago in evolutionary time—life in the Western world was very different from life now. Oliver Twist asked for more, and the food he wanted was gruel. Travel on foot or on horseback required the expenditure of energy. Occupation, often on the land, meant hard physical labor. Infectious diseases were prevalent. Despite the availability of epidemiologic methods we might now consider crude, society recognized the need for clean water to prevent cholera and, through appropriate intervention, reduced the risk of waterborne disease in general. In the 21st century, chronic diseases are prevalent. Many of these diseases are linked to obesity. The challenges for modern epidemiologists are to develop and use study designs and analytic methods capable of revealing the complex relationships involved in the etiology and consequences of obesity, as well as to develop and evaluate the individual and societal changes needed to reduce risk. If these challenges are met, the impact on health is likely to be large.
ABOUT THE AUTHOR
GAIL McKEOWN-EYSSEN is a Professor in the Departments of Public Health Sciences and Nutritional Sciences at the University of Toronto. She proposed that features of the metabolic syndrome may account for relationships of diet and physical activity to colorectal cancer. Study of biologic pathways linking lifestyle and cancer led her to an interest in the statistical models needed for such studies.
1. Potter JD. Epidemiologic research in the face of an obesity epidemic. Epidemiology. 2006;17:124–127.
2. Hatcher L. A Step-by-Step Approach to Using the SAS System for Factor Analysis and Structural Equation Modelling. NC: SAS Institute Inc; 1994.
3. Ward H. The Interrelationships of Education, Income, Lifestyle Factors, and Adiposity in the Ontario Food Survey. Department of Nutritional Sciences, University of Toronto; 2005.
© 2006 Lippincott Williams & Wilkins, Inc.