Next Generation Missing Data Methods in HIV Research

摘要 这一建议的主要动机是我们参与博茨瓦纳组合 预防项目（BCPP）是一个正在进行的大规模人类免疫缺陷病毒 (HIV)在博茨瓦纳的30个社区进行了随机分组预防试验。如在 大多数艾滋病毒预防研究、艾滋病毒状况数据不完整以及不回应有关 性行为是这项研究目前面临的一个重要挑战，数据可能会丢失， 在个体之间随机地以复杂的模式传播。认识到现有的统计 缺失数据的方法在很大程度上不适合完全解决艾滋病毒中的这一重要问题 研究，我们建议开发下一代缺失数据方法进展顺利 超越了当前的识别和推理理论。具体而言，我们建议（1）制定一项 统一理论的识别汇集在理论的最新发展， 基于因果图的识别，具有来自统计的最近识别结果 文学这将使我们能够建立条件，在复杂的缺失数据设置 正如在BCPP中，人们可以解开潜在人群的特征，这些特征可能是 科学兴趣的特点，无反应过程不一定是科学的 兴趣;（2）建立在（1）开发相应的逆概率加权和双 BCPP中数据可能缺失的统计推断的稳健方法， 随机和复杂模式;（3）开发新的半参数插补方法， 仅仅依赖于无反应过程中编码的假设，从而允许完整的 BCPP中的数据分布不受插补过程的影响;（4）开发 方便用户的软件，以便利广泛使用目标1 - 3中制定的方法， 在广泛的模拟研究中应用并展示其良好的性能， 回答BCPP中主要感兴趣的科学问题。

Abstract This proposal is largely motivated by our involvement with the Botswana Combination Prevention Project (BCPP) which is an on-going large scale human immunodeficiency virus (HIV) cluster randomized prevention trial conducted in 30 communities across Botswana. As in most HIV prevention studies, incomplete data on HIV status and nonresponse to queries about sexual behavior is an important challenge the study currently faces, with data likely missing not at random and in complex patterns across individuals. Recognizing that existing statistical methods for missing data are largely ill-suited to fully address this important problem in HIV research, we propose to develop the next generation of missing data methods going well beyond current theory of identification and inference. Specifically, we propose (1) to develop a unified theory of identification bringing together recent developments in the theory of identification based on causal graphs with recent identification results from the statistics literature. This will allow us to establish conditions under which in complex missing data settings as in the BCPP, one can untangle features of the underlying population which may be of scientific interest from features of the non-response process not necessarily of scientific interest;(2) to build on (1) to develop corresponding inverse-probability-weighted and doubly robust methods for statistical inference in the BCPP where data are likely to be missing not at random and in complex patterns; (3) to develop novel semiparametric imputation methods that solely rely on assumptions encoded in the nonresponse process, thus allowing the complete data distribution in the BCPP to remain unscathed by the imputation process; (4) to develop user-friendly software to facilitate widespread use of the methods developed in Aims 1-3, and to apply and demonstrate their good performance in extensive simulation studies as well as in answering scientific queries of primary interest in the BCPP.