Methods to find and model predictors of the causal effect of HAART

描述（申请人提供）：本项目开发并应用新的因果推断统计方法来调查：（1）感染后开始高活性抗逆转录病毒治疗（HAART）的时间对HIV-1感染开始阶段的短期和长期免疫重建的影响，（2）一旦CD 4计数下降到某一阈值以下，可能取决于病毒载量，（3）HAART开始后1年免疫激活和CD 4计数对长期有益临床结果的影响，所有患者均保持抑制性HAART治疗。前两个目标将告知何时启动HAART的临床决策过程。延迟HAART的启动具有延缓HAART的不良反应和耐药性的优点，因此可能改善长期生存。然而，延迟启动可能导致不可逆的免疫系统损伤。公共卫生中的这一重大问题在艾滋病毒感染的早期阶段尤其重要，因为目前还没有明确的治疗指南，考虑到目前疾病预防控制中心鼓励对艾滋病毒感染进行早期诊断的努力，这一问题就更重要了。第三个目标很重要，因为它可以帮助决定除了HAART之外的新疗法的目标，并为这些新疗法的效果研究提供信息。 对于所有这三个目标，随机临床试验可能是不道德的，因为它们将迫使患者在基线后的较长时间内继续停止或接受治疗。因此，该领域必须将这些治疗策略建立在观察数据分析的基础上，其中治疗决策不是随机的，病情较重的患者更有可能更早开始治疗，更晚停止治疗。因此，估计方法必须充分调整患者的特征，混淆了开始治疗的时间和关注的结果之间的关系。该提案概述了基于一类新的结构嵌套均值模型和一种新型的截尾权重逆概率的新因果推断方法，以调整这种指示混淆，因为以前的方法不直接适用于我们的设置。 这些新的方法产生了一个大类的估计量，和天真的估计量的选择是低效的，因为它们会导致大的标准误差。因此，使用这些天真的选择，无法在合理大小的样本中很好地估计具有编码治疗效应对当前协变量的依赖性的许多参数的模型。这一建议解决了这一方法及其应用中许多尚未解决的问题。这项建议有三个目标。第一个目标是开发准确和精确的估计，使用半参数模型的理论，提交给方法学期刊。第二个目标是应用这些方法，其结果将提交给有关艾滋病毒/艾滋病的期刊。第三个目标是将最终的软件公开提供给艾滋病毒/艾滋病和其他领域的研究人员。这些方法可能对其他研究产生很大影响，因为治疗适应症的时间依赖性混杂是一个非常常见的问题，在缺乏适当的统计方法（如本文提出的方法）的情况下会导致偏倚。

DESCRIPTION (provided by applicant): This project develops and applies new causal inference statistical methodology to investigate: (1) the impact of the time of initiation of Highl Active Anti-Retroviral Treatment (HAART) following infection on short- and long-term immune reconstruction in the beginning stages of HIV-1 infection, (2) the impact of treatment regimes that initiate HAART once CD4 count drops below a certain threshold, possibly depending on viral load, and (3) the impact of immune activation and CD4 counts one year since HAART initiation on long-term beneficial clinical outcomes had all patients remained on suppressive HAART. The first two aims will inform the clinical decision process of when to initiate HAART. Delaying HAART initiation has the advantage of postponing the adverse toxicities from HAART and drug resistance, and hence might improve long-term prognoses. However, delaying initiation can lead to irreversible immune system damage. This major issue in public health is especially relevant in the early phases of HIV-infection, for which no firm treatment guidelines exist, and even more so in view of the current CDC efforts to encourage early diagnosis of HIV-infection. The third aim is important because it can help decide whom to target with new therapies besides HAART and inform studies of the effect of such new therapies. For all three aims, randomized clinical trials may be unethical, since they would entail forcing patients to continue being off or on treatment over prolonged periods of time after baseline. Therefore, the field has to base these treatment strategies on analyses of observational data, where the treatment decisions are not randomized, and sicker patients are more likely to initiate treatment earlier and discontinue treatment later. Thus, estimation methods must adequately adjust for the patient characteristics that confound the relationship between times of initiating treatment and the outcome of interest. This proposal outlines new causal inference methods based on a new class of Structural Nested Mean Models and a new type of Inverse Probability of Censoring Weighting to adjust for this confounding by indication, since previous methods are not directly applicable to our settings. These new methods generate a large class of estimators, and naive choices of estimators are inefficient in that they lead to large standard errors. Thus, with those naive choices, models with many parameters encoding the dependence of the treatment effect on current covariates cannot well be estimated in samples of reasonable size. This proposal addresses many unresolved issues in this methodology and its applications. The objectives of this proposal are three-fold. The first objective is to develop accurate and precise estimators, using theory about semiparametric models, to be submitted to methodological journals. The second objective is to apply these methods, the results of which will be submitted to journals about HIV/AIDS. And the third objective is to make the resulting software publicly available for researchers in HIV/AIDS and other fields. These methods may have a great impact on other investigations, since time-dependent confounding by treatment indication is a very common problem that leads to bias in the absence of proper statistical methods like the ones proposed here.