Childhood overweight and obesity has emerged as an epidemic. In 2003-2004, over 33% of US children and adolescents were overweight or obese, and over 17% were obese. As the cohort born since 1980 moves into adulthood and middle age, we will see increasing incidence of diabetes, heart disease, kidney failure, and related metabolic disorders. Considerable research effort has been expended to identify the causes of childhood obesity, a necessary first step in suggesting potential prevention strategies to mitigate or reverse this growing problem. However, to date only particular subsets of the problem have been addressed.

Some of the identified factors related to obesity are clearly nested within others. For example, characteristics of the built environment vary by race-ethnicity, and obesigenic factors are more prevalent in disadvantaged communities; this may increase the risk of obesity and low physical activity levels in children. Likewise, metabolic processes are nested within individuals, who are further located within neighborhoods; as a result, insulin resistance may also follow proximity to healthy neighborhoods and their amenities.

Disentangling the web of causation of childhood obesity is a formidable task, and both available data and standard epidemiologic analyses may be inadequate to capture multiple and interacting levels of causation. Furthermore, the presence of feedback loops precludes standard approaches to causality, which are based on acyclicity and SUTVA assumptions. For example, obese children are more likely to be socially isolated, spending more time watching television, and less time engaged in sports and clubs than non-obese children. Causality is difficult to assess in this situation: obese children may be shunned by peers, making them more likely to spend isolated recreation time watching television, and less likely to participate in sports and clubs. Alternatively, participation in clubs and sports may prevent obesity by giving children opportunities for social interaction and physical activity.

In response to RFA-HD-08-023 (Innovative Computational and Statistical Methodologies for the Design and Analysis of Multilevel Studies on Childhood Obesity), we propose a novel multilevel study design and analysis plan to address this problem. In particular, we propose a three-stage strategy based on simulation, confirmation, and manipulation.

We propose the following specific aims:

1. Generate a series of simulated datasets based on knowledge obtained from the literature.

The simulations will be created using both a statistical simulation approache and an agent-based modeling approach, whereby social interactions follow explicit rules. The characteristics [heterogeneity] of individuals and environments will be specified using existing literature. We will focus on the following levels:

(1) Proximal determinants of obesity, which are in feedback and through which other determinants operate.

* Energy expenditure, which is a weakly correlated, negative, and linear determinant of adiposity

* Energy intake, which is a strongly correlated, positive, and nonlinear determinant of adiposity

* Energy expenditure and intake jointly determine BMI through complex feedback

(2) Distal demographic determinants of energy intake and expenditure:

* Household income, parental education level, gender, and race-ethnicity

(3) Family/household behavioral and biologic characteristics of energy intake and expenditure.

* Preparing meals at home; eating meals together; parental adiposity; genetic factors (e.g., FTO)

(4) Social network characteristics as they relate to energy intake and expenditure

* Homophily: on race, gender, parental education, adiposity; centrality, prestige

* Network formation: p* models for obesity as an explanation of friendship; sport/club affiliation

(5) Characteristics of neighborhood

* SES, school funding, and the built environment (supermarkets, fast food, and recreation outlets)

Characteristics 3 through 5 will further incorporate rules of social interaction between heterogeneous agents with each other and with the social and build environments. Institutions may themselves be agents. In order to be realistic, the construction of these simulated datasets will require a thorough understanding of the mechanisms and processes by which obesity and obesegenic environments are produced, not merely an ability to predict the shape of the obesity outcome.

2. Test the agreement between the simulated models and actual (observed) patterns of obesity.

Tests will be against available data, based on the characteristics cited in Aim 1, as well as overall U.S. and regional overweight and obesity rates. Data will come from the literature as well as representative datasets (NHANES, Add Health, ALSPAC). Tests must be limited to what data are available; ALSPAC will be used primarily to understand the longitudinal dynamics of obesity, while Add Health will focus on the role of social networks. Agreement between the regression-based and agent-based models may indicate that agent-based models are unnecessary.

3. Perform pseudo-experiments using the simulated data that manipulate individual characteristics and/or rules of agent interaction.

The true significance of simulated data comes not only from demonstrating how properly specified properties and rules create realistic patterns of obesity, but in the ability to artificially manipulate these properties and rules to provide counterfactual contrasts. We may then ask questions that are impossible to answer (or would yield implausible results) using the traditional multilevel approach, e.g.: what would be the effect on childhood eating habits of adding a supermarket to a "food desert", or eliminating the stigmatizing effects of overweight on friendship formation and club participation?

A related approach involves the application of structural equation models to the simulated and longitudinal datasets (ALSPAC, Add Health). [Palloni's Monte Carlo method for SEM]

We note that only existing ALSPAC data will be used for the proposed work; there is no need for further data collection.