Analytic Methods for Heterogeneous Multilevel Data

描述（申请人提供）：多水平数据在社会学、行为学和生物医学研究中非常常见。这些数据可以来自纵向社区调查、遗传家族研究或时空研究，以调查某些健康结果。通常情况下，兴趣集中在一些治疗干预的影响。当存在多级数据结构时，这样的数据可能非常复杂。数据可能包含社区、家庭、患者和重复测量等因素，这些因素随着时间的推移相互嵌套或交叉。对于连续响应，已研究并应用了分层模型，如线性混合效应模型或潜变量模型。在分析中，主要的兴趣是研究特定的原因路径对健康结果的影响。由于每个聚类中的记录通常是相关的，研究者必须调整观察聚类内或聚类之间的异质性。过度分散在这类数据中也很常见。本计画的主要目的是探讨上述性质之连续与离散结果的分析方法。在这一领域，人们通常将广义线性混合效应模型GLMM、边际模型或过渡模型应用于非连续数据。GLMM等模型的难点在于估计方法往往难以达到无偏性、一致性和有效性。我们感兴趣的是发展更强大的方法来实现这些目标的连续和离散的多级数据与任意尺寸。最终的结果是一个软件库，具有灵活的多级建模方法，用于分析复杂的多级数据。该软件将是有用的生物医学研究人员工作的社会学，行为和生物医学研究与复杂的数据结构。将编制手册和成套课程，以协助从业人员将适当的方法和软件工具应用于其研究。

DESCRIPTION (provided by applicant): Multilevel data are very common in sociological, behavioral and biomedical researches. The data could come from longitudinal community surveys, genetic family studies or spatial-temporal studies to investigate some health outcomes. Typically, the interest focuses on the impact of some treatment intervention. Such data could be very complex when there are multiple levels of data structures. The data might have factors such as community, family, patient and repeated measures over time nested or crossed in each other. For continuous response, hierarchical models such as linear mixed-effects models or latent variable models have been studied and applied. In the analysis, the major interest is to study the impact of specific cause pathway on health outcome. Since the records in each cluster are often correlated, investigator has to adjust the heterogeneity within a cluster of observations or between clusters. Overdispersion is also very common in such data. The major interest of this project is to investigate the analytic methods for continuous and discrete outcomes of the above nature. In this area, typically, people apply generalized linear mixed-effects models GLMM, marginal models or transition models to non-continuous data. The difficulties for such models such as GLMM is that estimation methods often have troubles to achieve unbiasness, consistency and efficiency. We are interested in the development of more robust methods to achieve these goals for continuous and discrete multilevel data with arbitrary dimension. The final result is a software library with flexible multilevel modeling approaches for the analysis of complex multilevel data. The software will be useful to biomedical researchers working on sociological, behavioral and biomedical studies with complex data structures. Manuscripts and course packs will be developed to assist practitioners in applying appropriate methods and the software tool to their studies.