Statistical Analysis of Longitudinal Studies and Surveys with Missing Values

-----------------------------------------------------------------------建议书编号：DMS 9803720 PI：罗德里克小机构：密歇根大学项目：纵向研究和缺失调查的统计分析摘要：许多实证研究都存在数据缺失的问题，其中缺少值的受试者与具有完整数据的受试者有系统地不同。丢弃不完整的案例会导致偏颇和效率的损失。基于模型的统计方法允许根据所有数据进行有效和有效的推断，但需要仔细注意数据的正确建模和丢失数据的机制。这项研究将开发新的贝叶斯和最大似然方法来处理以下问题中的不完整数据：纵向数据，研究对象随着时间的推移被重复测量，研究对象在特定时间丢失，或者过早退出研究；以及来自复杂调查设计的调查数据，其中研究对象具有不同的选择和不回答的概率。由于缺少数据，许多研究的结果很难解释。在民意调查中，一些随机选择的受试者拒绝回答一些问题，基于受访者的结果可能不能代表总体。在临床试验中，一些患者可能会因为无法耐受特定的治疗而退出研究，或者干脆搬到不同的地方而无法追踪。分析缺失值数据的统计方法使用非受访者的部分信息，以提供比仅分析完整案例更准确和更准确的答案。

----------------------------------------------------------------------- Proposal Number: DMS 9803720 PI: Roderick Little Institution: University of Michigan Project: Statisitcal Analysis of Longitudinal Studies and Surveys with Missing Values Abstract: Many empirical studies have problems with missing data, where subjects with missing values differ systematically from subjects with complete data. Discarding the incomplete cases leads to bias and loss of efficiency. Model-based statistical methods allow for efficient and valid inferences based on all the data, but require careful attention to correct modeling of the data and the missing data mechanism. This research will develop new Bayesian and maximum likelihood methods for handling incomplete data in the following problems: Longitudinal data where study subjects are measured repeatedly over time, and subjects are missing at certain times, or drop out of the study prematurely; and survey data from complex survey designs where subjects have differential probabilities of selection and nonresponse. Results from many research studies are hard to interpret because of missing data. In an opinion poll some of the randomly chosen subjects refuse to answer some of the questions, and results based on the respondents may not be representative of the population. In clinical trials some patients may drop out of the study because they cannot tolerate a particular treatment, or simply move to a different location and cannot be traced. Statistical methods for analyzing data with missing values use partial information on nonrespondents to provide more precise and more accurate answers than can be obtained by analyzing only the complete cases.