FAI: causal and semi-parametric inference for explanations of disparities and disparity-correcting modeling

随着学习算法在我们的生活中变得无处不在，许多人表达了对这些算法使用数据中的敏感特征-如种族、年龄、性别或健康-时可能出现的潜在有害偏见和差异的担忧。该项目旨在使用因果推理来理解和纠正这些差异，其目的是使用数据来量化因果关系。这种关系可以通过尝试预测当原因变量采用与其通常获得的值不同的值时影响的变化来揭示。例如，健康结果的种族差异可能是由于不同种族群体先前存在的共病比率不同，或者是由于隐性偏见引起的护理差异，或者是数据中没有衡量的一些其他机制造成的。群体依赖的合并症发生率对健康结局的间接影响可以概念化如下。测量A族患者的健康结果，然后使用可靠的模型预测相同患者的结果，人为地将共病比率设置为B族患者的结果，同时保持其他所有因素不变，并对两者进行比较。测量的健康结果和预测的健康结果之间的差异表明，不同的合并症比率是原因。同样，通过将A族患者的健康结果与同一组中预测的健康结果进行比较，可以揭示直接影响，其中所有参与造成差异的已知间接机制的变量保持不变，但表明族裔的变量变为B组。这种直接影响可被视为族裔对结果的总体影响的比例，不能用间接影响解释。本项目旨在开发使用数据来预测这种假设比较结果的方法，使用结果更好地理解医疗保健差异的机制，并建立能够意识到并可以纠正不良差异机制的预测模型。在这个项目中，研究人员旨在解决通过因果机制解释差异的概念、方法和实践方面的差距，并建立可以纠正被认为是不允许的机制的模型和工具。这种机制的一个例子是决定或结果直接依赖于感知到的种族或性别。研究人员将开发一种方法，评估在多大程度上可以将敏感特征的结果差异归因于不同的因果路径。为了解决在复杂的高维环境中因果推理面临的挑战，调查者将采用现代半参数方法，这些方法能够使用机器学习模型，同时保持期望的稳健性和快速收敛速度。此外，研究人员将开发因果和半参数推理以及约束优化的新方法组合，以创建预测模型和决策支持工具，这些模型和决策支持工具在有效使用数据的同时防止不允许的差异机制运行，例如通过确保感知到的种族对结果或决策没有直接影响。该方法将应用于量化差异，并使用从约翰·霍普金斯大学电子健康记录中获得的复杂数据集来建立决策支持工具。所有方法都将在一个开源软件包中实施，Ananke。这一奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

As learning algorithms become ubiquitous in our lives, many have expressed concerns about the potentially harmful biases and disparities that may arise when these algorithms use sensitive features in the data --- such as race, age, gender, or health --- inappropriately. This project aims to understand and correct for these disparities using causal inference, which aims to use data to quantify cause--effect relationships. Such relationships can be uncovered by trying to predict the change in an effect when a cause variable takes on a different value than one it usually attains. For example, ethnic disparities in health outcomes may arise from different rates of pre-existing comorbidities in different ethnic groups, or from differences in care arising from implicit biases, or from some other mechanisms unmeasured in the data. The indirect effect on health outcome of group-dependent rates of comorbidities can be conceptualized as follows. Measure the health outcomes in patients from ethnic Group A, then use a reliable model to predict outcomes for the same patients with comorbidity rates artificially set to that of ethnic Group B, while leaving everything else the same, and compare the two. Disparity between the measured and predicted health outcomes point to differing comorbidity rates as the cause. Similarly, a direct effect could be revealed by comparing health outcomes in patients from ethnic Group A with predicted health outcomes in that same group with all variables participating in known indirect mechanisms giving rise to disparities left intact, but the variable indicating ethnicity changed to that for Group B. Such a direct effect may be viewed as the proportion of the overall effect of ethnicity on the outcome not explained by indirect effects. This project aims to develop methods that use data to predict how such hypothetical comparisons would turn out, use the results to better understand mechanisms of disparities in healthcare, and build predictive models that are aware, and can correct for undesirable mechanisms of disparity.In this project, the investigator aims to address conceptual, methodological, and practical gaps in explaining disparities via their causal mechanisms and building models and tools that can correct for mechanisms deemed impermissible. An example of such a mechanism is a direct dependence of a decision or outcome on perceived race or gender. The investigator will develop methods that can assess the extent to which disparities in outcomes with respect to a sensitive feature can be attributed to distinct causal pathways. To address challenges causal inference faces in complex high-dimensional settings, the investigator will adopt modern semi-parametric methods that are able to use machine learning models while retaining desirable properties of robustness and rapids rates of convergence. In addition, the investigator will develop a novel combination of methods from causal and semi-parametric inference, and constrained optimization to create predictive models and decision support tools that use data efficiently while preventing impermissible mechanisms of disparity from operating, for instance by ensuring that perceived ethnicity has no direct effect on outcomes or decisions made. The approach will be applied to quantifying disparities and building decision support tools using a complex data set obtained from electronic health records at Johns Hopkins University. All methods will be implemented in an open-source software package Ananke.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.