Causal Inference with Irregularly Spaced Observation Times

点击翻译按钮获取中文摘要

This research project will develop a causal inference and machine learning toolset to tackle important and recurring challenges arising from emergent real-world data. Real-world data (e.g., consumer expenditures, mobile health applications, and electronic health records) provide unique opportunities for discovering optimal treatment strategies for the economy and health care. However, complex data also present novel challenges for statistical analysis. These challenges, such as irregularly spaced observation times or mixed data types, are impediments to effectively translating rich information into meaningful knowledge. This project will result in fundamental, broadly applicable advances in methodology for causal models with complex structures. It will provide principled causal inference approaches to scientific questions with complex data, such as longitudinal observational data, mobile health data, and electronic health records. The results of this research will be incorporated into graduate teaching, short courses, and workshops. Open-source software and R packages also will be developed.This research project will develop simple-to-interpret Marginal Structural Models for multinomial choices, taking into account correlations of expenditure categories, with an application to study the effect of lockdowns on consumer shopping behavior during the COVID-19 pandemic. Semiparametric doubly robust estimators will be developed to address time-varying confounding and irregularly spaced observation times, capitalizing on semiparametric efficiency theory and advanced machine learning methods. The investigator also will develop a unified framework of continuous-time Structural-Nested Models (SNMs) for general outcomes with time-varying confounding and informative observation times. The informativeness of observation times presents vital obstacles to the identification and estimation of the SNM parameter. Finally, electronic health records collect large amounts of granular patient data, which provide both opportunities and challenges for improving the assessment of treatment effects. The investigator will develop causal inference methods for estimating treatment effects with new functional principal component analysis (FPCA) of functional confounders subject to informative sampling for observations. The new FPCA also presents new prospects in the scope of functional data analysis.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.