Causal Inference and Machine Learning Methods

因果推理和机器学习方法

• 批准号: 1941419
• 负责人: Rajarshi Mukherjee
• 金额: $ 12.07万
• 依托单位: Harvard University
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2020
• 资助国家: 美国
• 起止时间: 2020-02-01 至 2022-01-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1941419&HistoricalAwards=false
• 关键词: Causal; Inference; Machine; Learning; Methods

Causal Inference is a broad area of statistical research where investigators are interested in the quantitative exploration of cause and effect relationships between exposures and outcomes. Questions that fall under this framework range across a vast canvass of applications, including, medical sciences and biology, economics, social sciences, and environmental health. Specific examples can include understanding the efficacy of treatments for a disease, the importance of genes and proteins on deciding biological functions, the interplay between environmental factors and genetic variations on human mortality, and the role of careful product placements to modulate market behaviors. This immense breadth of the statistical paradigm naturally comes with its share of subtleties and pitfalls. One of the major challenges in a statistically principled analysis of cause and effects is the presence of other factors, known as confounders, which often mislead investigators in falsely believing spurious relationships. Accounting for such confounders, therefore, becomes of great importance. The increasing ability of human beings to collect more and more data has made measurements of many such factors possible in common examples of causal inference studies. Although in principle, this has made causal inference a more feasible and exciting field, the mathematical formalism of such studies still come with a burden of assumptions to deal with these confounders -- which can often be extremely restrictive for practical applications. It is widely believed that tools from machine learning and artificial intelligence are natural choices to alleviate the burden of these assumptions. This project is aimed at understanding the role of these tools in disentangling causal inference related questions in a statistically principled and mathematically sound manner. As mentioned above, it is often argued that the use of machine-learning methods to nonparametrically estimate nuisance parameters alleviates the burden of the assumptions made in observational studies. Although true at heart, most machine learning methods are geared to attain low prediction errors in regression type problems -- whereas estimation of quantities like causal effects might require a somewhat different understanding. This project is, therefore, aimed at disentangling some machine learning algorithms used in the study of causal effects. The major goals of this project can be divided into the following regimes -- (i) exploring the crucial, and often overlooked, need of formal statistical methods for inferring causal effects which are adaptive over standard assumptions made in practice, (ii) a causal mediation analysis framework which paves the way for seamless application of state of the art machine learning methods, and (iii) the mathematical exploration of machine learning algorithms such as deep neural networks and generative adversarial networks in the context of these inferential problems. The developed understanding will be used to explore the effect of early life exposure to metal mixtures (like lead, arsenic, and cadmium exposures through drinking water) on late-life neurological diseases (such as Alzheimer's disease) and the potential role of high dimensional biomarkers such as EV miRNA's that might modulate such effects.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

因果推断是统计研究的广泛领域，研究人员对暴露与结果之间因果关系的定量探索感兴趣。在该框架之下的问题范围包括大量应用程序，包括医学科学和生物学，经济学，社会科学和环境健康。 具体的例子包括了解疾病治疗的疗效，基因和蛋白质对决定生物学功能的重要性，环境因素之间的相互作用以及人类死亡率的遗传变异以及仔细的产品放置在调节市场行为方面的作用。统计范式的巨大广度自然带有其微妙之处和陷阱。对因果关系的统计原则分析中的主要挑战之一是存在其他因素（称为混杂因素），这些因素常常误导了错误地相信虚假关系的研究人员。因此，考虑到这种混杂因素变得非常重要。人类收集越来越多的数据的能力不断提高，已经在因果推理研究的常见例子中测量了许多此类因素。尽管原则上，这使因果推论变得更加可行，更令人兴奋，但此类研究的数学形式主义仍然带来与这些混杂因素打交道的负担 - 对于实际应用通常可以极为限制。人们普遍认为，机器学习和人工智能的工具是减轻这些假设负担的自然选择。该项目旨在理解这些工具在以统计原则和数学上合理的方式解散因果推理问题中的作用。如上所述，通常认为使用机器学习方法非参数估计滋扰参数减轻了观察性研究中假设的负担。尽管本质上是真实的，但大多数机器学习方法都旨在达到回归类型问题中的低预测错误 - 而对因果效应等数量的估计可能需要有所不同的理解。因此，该项目旨在解散因果效应研究中使用的一些机器学习算法。该项目的主要目标可以分为以下制度 - （i）探索对推断因果效应的正式统计方法的至关重要，经常被忽视的，这些方法对实践中的标准假设具有适应性，（ii）中介分析框架为无缝应用机器学习方法的无缝应用铺平了道路，以及（iii）在这些推论问题的背景下，对机器学习算法（例如深神经网络和生成的对抗网络）进行数学探索。发达的理解将用于探索早期寿命暴露于金属混合物（例如铅，砷和通过饮用水暴露）对晚期神经系统疾病（例如阿尔茨海默氏病）的影响以及高尺寸生物标志物的潜在作用例如可能调节这种效果的EV Mirna。该奖项反映了NSF的法定任务，并通过使用基金会的知识分子优点和更广泛的影响审查标准来评估值得支持。

项目成果

On Nearly Assumption-Free Tests of Nominal Confidence Interval Coverage for Causal Parameters Estimated by Machine Learning

• DOI: 10.1214/20-sts786
• 发表时间: 2020-08-01
• 期刊: STATISTICAL SCIENCE
• 影响因子: 5.7
• 作者: Liu, Lin;Mukherjee, Rajarshi;Robins, James M.
• 通讯作者: Robins, James M.