Targeted Empirical Super Learning in HIV Research

描述（由申请人提供）：该项目的目的是研究和扩展一种称为目标经验学习的一般统计方法，其中包括最近开发的目标最大似然方法。这种新的、统一的统计学习方法的基本理论基础已经发展起来，我们建议将目标经验学习扩展为一种实用的产品，可以应用于紧迫的科学问题。在与临床艾滋病研究领域的主要科学家长期合作的基础上，我们将使用这种新颖的方法来解决有关艾滋病毒的研究问题。给定由n个独立且同分布的随机变量组成的观测数据，目标经验学习采用以下要素：i)定义感兴趣的参数；2)对感兴趣的参数进行建模，不指定干扰参数或只包含真正已知的建模假设；3)开发目标参数的鲁棒和高效（最大似然）估计器。该方法依赖于统一的交叉验证，在竞争性估计器之间进行选择，例如，筛选参数化，算法和/或降维（特别是对于讨厌的参数）的选择。重要的是，所采用的交叉验证准则评估了这些候选估计器相对于感兴趣的参数的性能。需要解决的具体应用包括：i)为治疗艾滋病毒感染患者开发最佳个体化治疗规则的模型和相应的有针对性的经验学习器；2)估计艾滋病毒突变的可变重要性/因果效应措施，以预测对药物组合的临床反应；3)估计依从性概况对hiv感染患者病毒学抑制的因果影响。我们将进一步开发和应用一种新的基于重采样的多重测试方法，以适当地解决我们对许多感兴趣的科学参数的同时测试和估计。

DESCRIPTION (provided by applicant): The aim of this project is to study and extend a general statistical methodology, called Targeted Empirical Learning, which includes a recently developed Targeted Maximum Likelihood methodology. The fundamental theoretical underpinnings of this new and unified approach to statistical learning have been developed and we propose to expand Targeted Empirical Learning into a practical product that can be applied to pressing scientific questions. Building on long-standing collaborations with leading scientists in the areas of clinical AIDS research, we will use this novel methodology to address research questions concerning HIV. Given observed data consisting of a realization of n independently and identically distributed random variables, Targeted Empirical Learning employs the following elements: i) defining the parameter of interest; 2) modeling the parameter of interest, leaving the nuisance parameters unspecified or only including truly known modeling assumptions; 3) developing targeted robust and highly efficient (maximum likelihood) estimators of the parameter of interest. The methodology relies on unified cross- validation to choose between competitive estimators indexed by, for example, choices of sieves parameterizations, algorithms, and/or dimension reductions (in particular for the nuisance parameters). Importantly, the cross-validation criterion employed evaluates the performance of these candidate estimators with respect to the parameter of interest. Specific applications to be addressed include the following: i) develop models and corresponding targeted empirical learners of optimal individualized treatment rules for treating HIV-infected patients, 2) estimate measures of variable importance/causal effects for mutations in the HIV virus for predicting clinical response to drug combinations; 3) estimate causal effects of adherence profiles on virologic suppression for HIV-infected patients. We will further develop and apply a novel resampling-based multiple testing methodology to properly address our simultaneous testing and estimation of many scientific parameters of interest.