Dynamic Multichain Graphical Models for the Analysis of Childhood Obesity Data

描述（由申请人提供）：本项目旨在制定有针对性的策略和方法，以解决儿童肥胖数据分析中具有挑战性的方法学问题。特别令人感兴趣的是统计和计算模型，这些模型可以关注被认为是儿童肥胖的直接或间接原因的多层次风险行为和风险因素。多层次视角可以通过行为-社会-生态概念模型来捕捉，其中个人因素，信仰；口味偏好;膳食成分;环境因素，如家庭、学校和食物供应；社会因素，如文化规范；而生理因素，如宫内和遗传倾向，要么作为解释儿童肥胖的因果链的因素，要么作为所谓的风险调节因子（一套稳定的生态条件），调节肥胖结果的概率。为了操作和实施概念多层模型，我们建议围绕我们称为动态多链图形模型（DMGM）的核心技术，建立一套相关的策略和方法，用于(1)数据预处理，以及(2)肥胖的多种因果途径建模。DMGM将直接风险因素和风险监管机构分为两个不同的空间，即所谓的因果空间和监管空间。对因果空间中基于机制的模型的兴趣集中在直接风险因素的联合分布上。或者，监管空间内的风险监管者通过对系统参数施加基于回归的模型来影响因果空间中的变量系统；然而，回归量的联合分布在这里没有多大意义。通过利用因果空间和监管空间分离所提供的概念和计算优势，DMGM能够处理三个或更多级别的数据，在一定程度上可以识别直接和间接风险变量。其他策略也可用于处理特定空间内的多级数据。除了DMGM，该项目还将包括一些其他工具的开发，这些工具专门用于解决儿童肥胖研究人员在实证工作中经常遇到的分析复杂性。该工具包包括一个可以处理时间数据的基于递归分区的决策树、一个用于处理历史数据的功能性数据分析工具，以及一个用于总结多个测量和处理空间内聚类效应的潜在变量模型。该项目的其他具体目标包括将拟议的方法应用于分别从路易斯安那州儿童健康研究和心跳！项目，并传播一个用户友好的软件程序，以增加项目的潜在影响。

DESCRIPTION (provided by applicant): This project aims to develop targeted strategies and methods for solving challenging methodological problems in the analysis of data on childhood obesity. Of special interest are statistical and computational models that can attend to the multiple levels of risk behaviors and risk factors that are deemed to be direct or indirect causes of childhood obesity. The multilevel perspective can be captured through a behavioral-social-ecological conceptual model in which personal factors, beliefs; taste preferences; dietary composition; environmental factors such as homes, schools, and food availability; societal factors such as cultural norms; and physiological factors such as intrauterine and genetic disposition are included either as factors in causal chains explaining childhood obesity or as so-called risk regulators (a set of stable ecological conditions) that up- and down-regulate probabilities of obesogenic outcomes. In order to operationalize and implement the conceptual multilevel model, we propose to build, around a core technology that we call the Dynamic Multi-chain Graphical Model (DMGM), a set of related strategies and methods for (1) the preprocessing of data, and (2) the modeling of multiple causal pathways to obesity. The DMGM separates direct risk factors and risk regulators into two distinct spaces the so-called causal space and the regulatory space. Interest in the mechanism-based model within the causal space focuses on the joint distribution of direct risk factors. Alternatively, risk regulators within the regulatory space affect the system of variables in the causal space through regression-based models imposed upon system parameters; the joint distribution of regressors, however, is of little interest here. By capitalizing on the conceptual and computational advantages offered by the segregation of the causal and regulatory spaces, the DMGM is able to handle three or more levels of data, to the extent to which direct and indirect risk variables can be identified. Other strategies are also available for handling multiple levels of data within a specific space. Besides the DMGM, this project will also include the development of a number of other tools that are especially designed to address the analytical complexity that childhood obesity researchers often encounter in their empirical work. The toolkit includes a recursive-partition-based decision tree that can handle temporal data, a functional data-analysis tool for processing history data, and latent-variable models for summarizing multiple measurements and handling within-space clustering effects. Other specific aims of the project include the application of the proposed methods to two national data sets collected, respectively, from the Louisiana Child Health Study and the Heartbeat! Project, and the dissemination of a user-friendly software program for increasing the potential impact of the project. PUBLIC HEALTH RELEVANCE: As a social epidemic, childhood obesity is the result of the interaction between many levels of personal behavior and risk factors, as well as obesogenic ecological factors that span many sources, including families, schools, and communities. Using a broad and interdisciplinary team of clinical and methodology scientists, this project develops advanced analytic tools that could help clinicians better understand the mechanisms of how multiple levels of risk factors lead to childhood obesity, including the relative importance of the risk factors.