Small Area Estimation for State and Local Health Departments

项目摘要 州和地方卫生部门的研究人员在进行小面积估计时往往面临双输的局面。 一方面，有大量证据表明，在许多健康结果及其风险因素方面存在种族差异， 但按空间和种族（以及年龄和性别等因素）对数据进行分层只会加剧问题 通过将具有小样本大小的数据集划分为较大的数据集， 样本量较小。另一方面，虽然可以使用复杂的统计模型来产生 从有限的数据中得出更精确的估计，州和地方卫生部门的估计可能是 被视为“官方统计”，因此这些机构可能不愿意过于依赖统计模型， 害怕他们可能带来的偏见。 拟议工作的目标有三个方面。我们的第一个任务将是开发统计模型， 多变量空间数据的分析，允许用户预先指定模型信息量的上限 - 也就是说，当生成基于模型的数据时，与数据相比，给予模型的权重的度量 估算这项工作将建立在丰富的空间统计文献和最近的研究，提供了深入了解 如何量化空间模型的信息量。我们将把这种方法扩展到多变量的设置， 空间数据，用于计算特定人口群体估计值和年龄调整估计值。 因为我们设想这些方法对州和地方卫生部门的研究人员有用， 我认为应该对我们的方法进行彻底的案例研究，以评估它们的适用性。为此，我们 第二项任务是与费城公共卫生部合作， 该研究旨在对费城的心脏病死亡率及其危险因素进行严格的分析。这 分析将产生每年一次的人口普查区域水平的估计，由于几种形式的心脏病死亡率 以及按年龄、性别和种族/民族分类的关键风险因素的患病率估计值。这项研究的成果 将包括一系列报告-一个侧重于城市一级的趋势，一个侧重于社区一级的趋势 趋势-一个互动的在线仪表板，以及同行评审的出版物，为我们的发现添加背景。 最后，我们认识到，很少有州和地方卫生部门的工作人员谁是先进的培训， 空间贝叶斯统计方法，这一事实可能会成为我们使用的方法的障碍， 开发.为了解决这个问题，我们的第三个任务是与CDC资助的GIS能力建设项目合作， 为州和地方卫生部门提供地理空间分析培训。为期一个月的培训计划 首先向用户介绍ArcGIS软件包，最后概述了由 地理信息系统能力建设项目-速率稳定工具。在这个项目中，我们将与GIS合作， 能力建设项目，将我们开发的方法纳入“黑箱”框架中， 提供关于在疾病监测中使用空间贝叶斯方法的额外培训。

PROJECT SUMMARY Researchers at state and local health departments producing small area estimates often face a lose-lose situation. On one hand, there is a wealth of evidence of racial disparities in many health outcomes and their risk factors, but stratifying data by space and race (in addition to factors such as age and sex) only exacerbates the issues associated with small area estimation by dividing a dataset with small sample sizes into a larger dataset with smaller sample sizes. On the other hand, while the use of complex statistical models can be used to produce more precise estimates from limited data, estimates produced by state and local health departments may be treated as “official statistics” and thus these agencies may be reluctant to rely too heavily on statistical models for fear of the bias they may introduce. The objective of the proposed work is three-fold. Our first task will be to develop statistical models for the analysis of multivariate spatial data that allow users to pre-specify an upper bound on the model's informativeness — i.e., a measure of the weight given to the model as compared to the data when producing model-based estimates. This work will build on the rich spatial statistics literature and recent research that provides insight into how to quantify the informativeness of spatial models. We will extend this approach to the setting of multivariate spatial data for the purposes of calculating demographic group-specific estimates and age-adjusted estimates. Because we envision these methods being useful for researchers at state and local health departments, we believe a thorough case study of our methods should be conducted to assess their suitability. To this end, our second task will be to partner with the Philadelphia Department of Public Health and use the methods we've developed to conduct a rigorous analysis of heart disease mortality and its risk factors in Philadelphia. This analysis will produce yearly census tract-level estimates for rates of death due to several forms of heart disease and estimates of the prevalence of key risk factors by age, gender, and race/ethnicity. The product of this research will include a collection of reports — one focused on city-level trends and one focused on neighborhood-level trends — an interactive online dashboard, and peer-reviewed publications that add context to our findings. Finally, we recognize that few state and local health departments have staff who are trained in advanced spatial Bayesian statistical methods, a fact that could serve as an impediment to the use of the methods we develop. To remedy this, our third task will be to partner with the CDC-funded GIS Capacity Building Project, which provides training in geospatial analyses to state and local health departments. This month-long training program begins by introducing users to the ArcGIS software package and concludes with an overview of a tool created by the GIS Capacity Building Project — the Rate Stabilizing Tool (RST). For this project, we will partner with the GIS Capacity Building Project to incorporate the methods we develop into the RST in a “black-box” framework and provide additional training on the use of spatial Bayesian methods in disease surveillance.