Hierarchical Bayesian Analysis of Complex Sample Survey Data

描述（申请人提供）：拟议的研究将使用分层贝叶斯模型来解决基于人口的调查数据分析中的三个相互关联的问题：由于样本设计或抽样后无响应而导致的不相等的纳入概率;项目级数据中不可重复的缺失;以及组合来自多个复杂调查数据集的信息以获得对人口数量的更准确和有效的估计。我们打算开发强大的模型，可以提供“数据驱动”的权重修剪程序的一般类别的人口统计数据下的各种样本设计;开发选择模型，以适应不可分割的缺失机制的背景下，复杂的调查设计;并制定方法，联合收割机数据从多个调查创建合成人口从每个调查，然后结合这些人口开发估计。虽然我们的方法将适用于各种各样的分析程序，但我们将特别关注小面积或小域估计，因为本提案旨在解决的问题通常在环境中最为严重。具有高度可变权重的域估计器可能具有较差的均方误差特性。不可解释的无应答和区域/域之间的关联可能会使域间比较不可靠。如果有正确的程序来考虑复杂的抽样设计以及可能出现的无答复偏差和计量误差，则一项调查中某一领域的小样本可以用其他调查的数据来补偿。我们将考虑三个主要应用：使用国家健康和营养检查调查进行分析，以确定儿童出生体重与心血管风险因素之间的关联，通过结合行为风险因素监测调查和国家健康访谈调查的数据，确定成人中癌症行为风险因素的患病率，并在美国改变生活小组调查中探索老年人的死亡率压缩。分析将侧重于小领域（例如种族/少数民族和县/州）。虽然该方法的动机是从贝叶斯的角度来看，结果将从基于设计的角度使用分析和模拟技术进行评估。我们亦会集中开发方便使用的软件，以推行新方法。公共卫生相关性：在一个日益多样化的国家，越来越需要有针对性地开展公共卫生研究，并向小地区提供服务，无论是地理上的还是人口上的（如少数民族）。健康调查是开展这些工作的丰富数据来源，但提取小地区信息的方法仍不完善。拟议的工作将制定新的方法来处理小面积估计造成的一些问题，包括由于样本规模小和选择概率不等而造成的不稳定估计，以及由于选择参加调查的人与拒绝或无法联系的人之间的差异而造成的有偏差估计。这项工作还可以通过开发新的方法来合并多项调查的数据，提高目前收集的数据的使用效率。

DESCRIPTION (provided by applicant): The proposed research will use hierarchical Bayesian modeling to tackle three interrelated problems in the analysis of population-based survey data: accounting for unequal probabilities of inclusion due to sample design or post-sampling non-response; accounting for non-ignorable missingness in item-level data; and combining information from multiple complex survey data sets to obtain more accurate and efficient estimates of the population quantities. We intend to develop robust models that can provide "data- driven" weight trimming procedures for a general class of population statistics under a variety of sample designs; develop selection models that accommodate non-ignorable missingness mechanisms in the context of complex survey designs; and develop methods to combine data from multiple surveys by creating synthetic populations from each survey and then combining these populations to develop estimates. While our methods will be applicable to a wide variety of analytic procedures, we will focus on small area or small domain estimation in particular, since the issues that this proposal intends to address are often most acute in the setting. Domain estimators with highly variable weights can have poor mean square error properties. Associations between nonignorable nonresponse and areas/domains can make between-domain comparisons unreliable. Small samples in a given domain in one survey can be compensated by data from other surveys, if correct procedures are in place to account for complex sample design, as well as the possibility of non-response bias and measurement error. We will consider three major applications: analyses to determine associations between birth weight and cardiovascular risk factors in children using the National Health and Nutrition Examination Survey, to determine the prevalence of cancer behavioral risk factors among adults by combining data from the Behavioral Risk Factor Surveillance Survey and the National Health Interview Survey, and to explore mortality compression among the elderly in the Americans Changing Lives panel survey. Analyses will focus on small domains (race/ethnic minorities, and counties/states, as examples). Though the method is motivated from a Bayesian perspective, the results will be evaluated from the design-based perspective using analytical and simulation techniques. We will also focus on developing user-friendly software to implement the new methods. PUBLIC HEALTH RELEVANCE: In an increasingly diverse nation, the need is increasing to target public health studies and delivery to small areas, be they geographic or demographic (such as ethnic minorities). Health surveys are a rich source of data for such efforts, but methods for extracting information about small areas remain undeveloped. The proposed work will develop new methods for dealing with some of the problems that small area estimation poses, including unstable estimates due to small sample sizes and unequal probabilities of selection, and biased estimates due to differences between people who chose to participate in the surveys and those who refused or could not be contacted. The work can also improve the efficient use of data currently collected by developing new ways of combining data from multiple surveys.