Hierarchical Models in Health Services Research

描述(由申请人提供)：提供者和医生概况、基因表达研究的数据结构和推断目标；大多数与健康相关的科学和政策研究需要一个分层的、随机的影响模型。有效和高效地估计总体参数、方差分量和单位特定的随机效应(提供者、临床医生、地区或基因特定的潜在属性)，所有这些都是在有效的不确定性评估的背景下进行的。适当开发的贝叶斯模型有效地实现了这些目标，考虑了方差分量和其他不确定性，改进了对随机影响的估计。在损失函数的指导下，该方法构建了非标准推理，例如排名(包括识别表现极差和表现良好的人)和估计随机单位特定效果的直方图。推论必须很好地实现目标目标(S)，并将效率(充分利用现有信息)与稳健性(在广泛的潜在假设范围内保持高效率)结合起来。拟议的研究建立在最近开发和评估的排名/百分位数方法的基础上。这些研究确定了需要进一步研究的领域，包括解决当前方法潜在缺陷的新排序方法、开发方法操作特征的数据分析量化、实施单变量和多变量稳健先验和直方图估计。研究将基于数学和基于模拟的分析；对卫生服务和基因表达数据集的分析，这些数据集说明了方法并产生了实质性的结果。继续开发基于R的模拟和数据分析软件将确保“可重复的研究”，并允许其他人实施程序。约翰霍普金斯大学的一名生物统计学博士前学生将得到支持和指导。拟议的研究将产生创新的统计方法，为解决多个目标的权衡提供指导，交流有趣和信息量大的案例研究，教育博士生并使其适应社会文化。

DESCRIPTION (provided by applicant):The data structure and inferential goals for provider and physician profiling, gene expression studies; most health-related science and policy studies require a hierarchical, random effects model. Valid and efficient estimation of population parameters, variance components, and unit-specific random effects (provider, clinician-, region- or gene-specific latent attributes), all in the context of valid uncertainty assessments. Properly developed Bayesian models effectively accomplish these goals, accounting for variance components and other uncertainties, improving estimation of random effects. Guided by a loss function, the approach structures non-standard inferences such as ranking (including identification of extremely poor and good performers) and estimating the histogram of random unit-specific effects. Inferences must perform well for the target goal(s) and combine efficiency (making good use of the available information) with robustness (high efficiency over a broad range of underlying assumptions). The proposed research builds on recently developed and evaluated ranking/percentiling methods. These studies have identified areas requiring additional research, including new ranking methods to address potential drawbacks of current methods, developing data analytic quantification of a method's operating characteristics, implementation of univariate and multivariate robust priors and histogram estimates. Research will be based on mathematical and simulation-based analysis; analysis of health services and gene expression datasets that illustrate approaches and produce substantive findings. Continued development of R-based, simulation and data analysis software will ensure "reproducible research" and allow others to implement procedures. A pre-doctoral Biostatistics student at Johns Hopkins will be supported and mentored. Proposed research will produce innovative statistical methodology, provide guidance on trade-offs in addressing multiple goals, communicate interesting and informative case studies, educate and acculturate a pre-doctoral student.