Targeted Learning: Causal Inference Methods for Implementation Science

有针对性的学习：实现科学的因果推理方法

• 批准号: 8900155
• 负责人: Mark J Vanderlaan
• 金额: $ 46.08万
• 依托单位: UNIVERSITY OF CALIFORNIA BERKELEY
• 依托单位国家: 美国
• 财政年份: 2007
• 资助国家: 美国
• 起止时间: 2007-07-01 至 2018-06-30
• 项目状态: 已结题
• 来源: https://reporter.nih.gov/project-details/8900155
• 关键词: AIDS prevention; Accounting; Acquired Immunodeficiency Syndrome; Address; Africa; Algorithms; Anti-Retroviral Agents; CD4 Lymphocyte Count; Caring; Clinic; Clinical; Cluster randomized trial; Collaborations; Communities; Community Health; Complex; Computer software; Data; Data Sources; Databases; Eastern Africa; Educational process of instructing; Enrollment; Epidemiology; Evidence based practice; Failure; Funding; Goals; Grant; HIV; Health; Individual; International; Intervention; Investigation; Joints; Kenya; Learning; Link; Measures; Mediating; Mediation; Methodology; Methods; Modeling; Monitor; Nurses; Observational Study; Outcome; Paper; Pathway interactions; Patient risk; Patients; Prevention strategy; Randomized; Recording of previous events; Regimen; Research; Research Design; Research Personnel; Resources; Sampling; Science; Series; Southern Africa; Specific qualifier value; Structural Models; System; Testing; Time; Triage; Uganda; Uncertainty; Work; Writing; antiretroviral therapy; base; causal model; cohort; comparative effectiveness; design; health care delivery; implementation science; improved; innovation; interest; mortality; novel; programs; research study; response; software development; theories; tool; treatment strategy

DESCRIPTION (provided by applicant): This competitive renewal will develop general methods for evaluating the comparative effectiveness of alter- native strategies for HIV prevention, treatment and care in Southern and Eastern Africa. Large cluster random- ized trials and global cohort collaborations generate longitudinal data on hundreds of thousands of patients in real world settings. These provide a tremendous resource for developing the "practice-based evidence" needed to maximize the impact of HIV prevention strategies and to improve healthcare delivery systems. Realizing this potential, however, demands innovations to the field of Targeted Learning for maximally unbiased and efficient estimation of statistical parameters, best approximating the causal effects of interest. First, improved methods for estimating the effects of patient responsive monitoring and treatment strategies must be developed. In the common settings of strong confounding and rare outcomes, current estimators suffer from bias, lack efficiency and have unreliable measures of uncertainty. Second, general causal models and identifiability assumptions must be developed for the joint effects of cluster and individual-level interventions over multiple time points. These models will account for interactions between individuals within clusters and potential contamination between clusters. This work will inform the optimal design for sampling clusters and measuring individuals within communities or clinics. Third, efficient and maximally unbiased estimators must be developed to evaluate the impact of cluster and individual-level interventions over multiple time points. Current methods are highly susceptible to bias and misleading inference due to model misspecification and due to the often incorrect assumption that the observed data represent n independent, identically distributed (i.i.d.) repetitions of an experiment. The developed methods will elucidate the pathways by which cluster- based interventions impact health, while remaining robust to the common challenges of sparsity, irregular and informative missingness, and few truly independent units (clusters) but potentially hundreds of thousands of conditionally independent units. These innovations are motivated by our collaborations with the International epidemiologic Databases to Evaluate AIDS (IeDEA) in Southern (PI Dr. Egger) and Eastern Africa (PI Dr. Yiannoutsos) and the Sustain- able East Africa Research in Community Health (SEARCH) consortium (PI Dr. Havlir), a cluster randomized trial to evaluate the community-wide benefits of ART initiation at all CD4 counts. The developed methods will be applied to these data sources to investigate (i) strategies for monitoring antiretroviral therapy (ART) and guiding switches to second line regimens, (ii) the direct and indirect effects of a community-based HIV prevention strategy and (iii) the impact of clinic-based programs for delivering HIV care. Finally, the resulting estimators will be implemented as publicly available software packages and teaching papers written to explain the methodology in a clear and rigorous manner.

描述（由申请人提供）：本次竞争性更新将制定一般方法，用于评估南部和东部非洲艾滋病毒预防、治疗和护理的本土策略的相对有效性。大型集群随机试验和全球队列合作产生了真实的世界背景下数十万患者的纵向数据。这些研究为开发“基于实践的证据”提供了巨大的资源，这些证据是最大限度地发挥艾滋病毒预防战略的影响和改善医疗保健提供系统所必需的。然而，实现这一潜力需要在目标学习领域进行创新，以最大限度地无偏和有效地估计统计参数，最好地近似感兴趣的因果效应。 首先，必须开发用于估计患者响应监测和治疗策略的效果的改进方法。在常见的强混杂和罕见的结果，目前的估计遭受偏见，缺乏效率和不可靠的措施的不确定性。其次，必须为集群和个人层面的干预措施在多个时间点的联合效应制定一般因果模型和可识别性假设。这些模型将考虑集群内个体之间的相互作用以及集群之间的潜在污染。这项工作将为抽样集群和测量社区或诊所内的个人的最佳设计提供信息。第三，必须开发有效和最大无偏估计，以评估多个时间点的集群和个人层面的干预措施的影响。目前的方法是非常容易受到偏见和误导性的推断，由于模型误指定，并由于经常不正确的假设，观察到的数据代表n独立，同分布（i.i.d.）重复一个实验。所开发的方法将阐明基于聚类的干预措施影响健康的途径，同时对稀疏、不规则和信息缺失以及很少有真正独立的单元（聚类）但可能有数十万个条件独立的单元的常见挑战保持稳健。 这些创新是由我们与南部非洲（PI Egger博士）和东非（PI Yiannoutsos博士）评估艾滋病的国际流行病学数据库（IeDEA）以及可持续的东非社区卫生研究（ESTA）联盟（PI Havlir博士）合作推动的，这是一项评估所有CD4计数的ART启动的社区范围内益处的群集随机试验。所开发的方法将应用于这些数据源，以调查（i）监测抗逆转录病毒治疗（ART）和指导转换到二线治疗方案的策略，（ii）基于社区的艾滋病毒预防策略的直接和间接影响，以及（iii）基于诊所的艾滋病毒护理方案的影响。最后，由此产生的估计将作为公开可用的软件包和教学论文，以明确和严格的方式解释的方法。