Transporting established insights from classical experimental design to address causal questions in environmental epidemiology including the understanding of biological mediating mechanisms

 描述(申请人提供)：由于缺少数据，缺乏随机性，以及由于时间性而导致的复杂性，在正确处理观察性研究中的因果关系方面存在根本差距。联系的衡量标准不适合提出相关的政策建议，因为这些建议涉及干预建议，这是因果陈述。长期目标是解决重要的环境健康因果问题(例如，发现将空气污染与低出生体重或自闭症联系起来的生物机制，将极端天气条件与中暑或全球营养缺乏联系起来)，并制定统计方法，正确处理环境健康研究中的因果关系。这项赠款申请的总体目标是通过传递过去80年(主要是自Fisher，1935年以来)在缺失数据领域(Rubin 1978)和经典和现代多因素随机实验中开发的成功的统计工具，正确地制定和估计因果环境健康影响，特别是在存在中间变量的情况下。在初步发展和重大应用的指导下，这项研究将包括四个具体目标：1)扩展成功的高维缺失数据多重填补方法，使其能够在面对缺失数据时使用标准的完全数据方法进行有效的统计推断。将考虑两种不同的设置，第一种是处理多变量时间序列，第二种是根据不太准确但可用的测量结果进行“黄金标准”预测。2)发展统计理论，从观察性研究收集的数据中估计临时预测，这些数据将被重建为来自随机实验的近似数据；一个特别有趣的设置将考虑暴露和结果之间因果路径上的中间变量(也称为“中介”)。3)扩展因果中介分析的标准方法；中介分析已成为检验因果生物学途径及其对不良健康影响的相对贡献的流行开发工具。4)使用与生物医学研究人员目前使用的软件兼容的新软件来实现在前三个目标中开发的这些方法。这项旨在正确表述和估计因果环境健康影响的拟议研究具有创新性，因为它通过在两个领域传递经典和现代统计学中发展的成功方法和概念而实质上偏离了现状：1)处理缺失数据的多重归因技术；2)复杂多因素随机试验的分析，特别是在存在中间变量和复杂数据(如纵向、生存和高维数据)的情况下。拟议的统计方法将对生物医学研究具有重要意义，因为它将在精确陈述的假设下产生有效的因果环境健康影响估计，预计将对政策决策产生积极影响，并建议适当的干预措施。

 DESCRIPTION (provided by applicant): There is a fundamental gap in correctly addressing causality in observational studies due to missing data, and lack of randomization, and complications due to temporality. Measures of association are inapposite for making relevant policy recommendations because these involve suggestions for interventions, which are causal statements. The long-term goal is to address important environmental health causal questions (e.g., discovering biological mechanisms linking air pollution to low birth weight or autism, and relating extreme weather conditions to heat stroke or worldwide nutritional deficiency) that have policy-relevant consequences, and develop statistical methodology that correctly addresses causality in environmental health studies. The overall objective of this grant application is to correctly formulate and estimate causal environmental health effects, especially in the presence of intermediate variables, by transporting successful statistical tools developed in the fields of missing data (Rubin 1978) and classical and modern multi-factorial randomized experiments over the past 80 years (essentially since Fisher, 1935). Guided by preliminary development and significant applications, this proposed research will consist of four specific aims: 1) Expand successful multiple- imputation methods for high-dimensional missing data to enable valid statistical inference when confronted with missing data using standard complete-data methods. Two different settings will be considered, the first one dealing with multivariate time series and the second with "gold standard" prediction from less accurate but available measurements. 2) Develop statistical theory to estimate casual estimands from data collected by observational studies, which would be reconstructed to approximate data from a randomized experiment; one particularly interesting setting will consider intermediate variables on the causal pathway between an exposure and an outcome (also called "mediators"). 3) Expand standard methods developed for causal mediation analysis; analysis of mediation has become a popular developing tool to examine causal biological pathways and their relative contribution to adverse health effects. 4) Implement these methods developed in the three previous aims with new software that is compatible with software currently used by biomedical researchers. The proposed research targeting correct formulation and estimation of causal environmental health effects is innovative because it represents a substantial departure from the status quo by transporting successful methods and concepts developed in classical and modern statistics in two areas: 1) multiple imputation techniques for handling missing data, 2) analysis of complex multi-factorial randomized experiments, especially in the presence of intermediate variables and complex data (e.g., longitudinal, survival, and high-dimensional). The proposed statistical methodology will be significant to biomedical research because it will yield valid causal environmental health effect estimates under precisely stated assumptions, which are expected to provide positive impacts on policy decisions and suggest appropriate interventions.