Design and Inference for Hybrid Ecological Studies

描述（由申请人提供）：生态学研究可以定义为在群体水平上检查关联。它们的吸引力在于它们利用了常规可用的数据，并且由于人口众多和广泛的暴露对比，它们还提供了高功率的潜力。然而，它们也容易受到个人层面关联的一系列偏见的影响，统称为生态偏见，并可能导致生态谬误。在流行病学中，最根本的困难是生态数据无法描述暴露和混杂因素的组内变异性。这导致无法控制混淆，以及个人层面模型的一般不可识别性。解决生态推断问题的唯一办法是用个体层面的样本来补充生态数据；在本提案中，我们描述并发展了各种追求这一解决方案的混合研究。具体来说，我们开发了一种混合设计，其中病例对照研究嵌入在生态研究中。直观的吸引力在于，个人层面的数据为控制偏见提供了基础，而生态数据提供了效率收益。此外，我们将现有的方法，包括聚合数据设计和两阶段法，扩展到生态环境。这将基于这些设计的贝叶斯方法的发展，这还没有被探索。此外，我们将比较各种方法在各种数据/采样场景中的性能。一个关键的研究问题是群体层面的数据是否为个人的收集提供了有用的信息。我们将探讨最佳的研究设计，从多少个人取样和从哪些群体。用两个癌症数据集和一个流感数据集说明了这些方法。

DESCRIPTION (provided by applicant): Ecological studies may be defined examining associations at the group level. They are appealing in that they make use of routinely available data, and also offer the potential of high power due to large populations and broad exposure contrasts. However, they are also susceptible to a range of biases with respect to individual-level associations, collectively termed ecological bias, and may lead to the ecological fallacy. In epidemiology, the fundamental difficulty is the inability of ecological data to characterize within-group variability in exposures and confounders. This results in an inability to control for confounding, and general non-identifiability of the individual-level model. The only solution to the ecological inference problem is to supplement ecological data with individual-level samples; in this proposal we describe and develop a variety of hybrid studies that pursue this solution. Specifically, we develop a hybrid design in which a case-control study is embedded within an ecological study. The intuitive appeal is that the individual-level data provide the basis for the control of bias, while the ecological data provide efficiency gains. In addition, we extend current methods, including the aggregate data design and two-phase method, to the ecological setting. This will be based on the development of Bayesian methods for these designs, which have not been explored. Further, we will compare performance of the various methods in a variety of data/sampling scenarios. A key research question is whether the group-level data provide useful information for the collection of individuals. We will explore optimal study design in terms of how many individuals to sample and from which groups. The methods are illustrated with two cancer data sets and one influenza data set.