Marginal Causal Models for Observational Studies in Obesity Research

描述（由申请人提供）：关于自然发生的行为（如节食）对超重和肥胖的影响有许多研究问题，随机对照试验是站不住脚的。对于这些，我们必须依靠观察性研究，而用观察数据进行因果推理在方法学上提出了困难的挑战。已经提出了许多估计平均因果效应的技术，但对这些效应的变化进行建模的工具很少。在这个R21项目中，我们将开发一种新的因果回归方法，允许个体水平的影响与个体、背景和环境的特征共变。我们的公式类似于James Robins等人开发的边际结构模型，但我们提出了一种基于imputation而不是加权的新估计技术。我们将获得估计和标准误差，诊断以帮助用户识别模型中的缺点，以及处理具有复杂设计特征（地层，集群，不平等选择概率）的调查数据的方法。方法将在用户友好的软件中实施，并提供给健康结果研究人员。将对三个流行病学数据集进行二次分析，以评估饮食限制、体育活动和其他行为对青少年和青年期体重和其他后遗症的影响变化。公共卫生相关性：该项目将(a)产生新的统计方法和软件，以帮助肥胖研究人员得出关于非随机行为和治疗效果的有力结论，(b)将这些方法应用于青少年和青年期体重相关行为的观察数据的二次分析。方法和研究结果将为设计和实施更有效的治疗和预防肥胖的干预措施提供信息。

DESCRIPTION (provided by applicant): There are many research questions about effects of naturally0ccuring behaviors (e.g. dieting) on overweight and obesity for which a randomized controlled trial would be untenable. For these we must rely on observational studies, and causal inference with observational data poses difficult methodological challenges. Many techniques have been proposed for estimating average causal effects, but tools for modeling variation in these effects are scarce. In this R21 project, we will develop a new methodology for causal regression which allows effects at the individual level to covary with characteristics of individuals, contexts and environments. Our formulation is similar to that of the marginal structural model developed by James Robins et al., but we propose a new estimation technique based on imputation rather than weighting. We will obtain estimates and standard errors, diagnostics to help users identify shortcomings in the model, and methods to handle data from surveys with complex design features (strata, clusters, unequal probabilities of selection). Methods will be implemented in a user-friendly software and made available to health-outcome researchers. Secondary analyses will be performed on three epidemiologic datasets to assess the variation in effects dietary restraint, physical activity and other behaviors on body weight and other sequelae in adolescence and young adulthood. PUBLIC HEALTH RELEVANCE: This project will (a) generate new statistical methods and software to help obesity researchers draw robust conclusions about effects of behaviors and treatments that have not been randomized, and (b) apply these methods in secondary analyses of observational data on weight-related behaviors in adolescence and young adulthood. Methods and findings will inform the design and implementation of more effective interventions for treatment and prevention of obesity.