Thinking Causally About High-Dimensional Databases
Wednesday, November 7, 2018 - 11:30am - 12:00pm
A movement that is convergent across many academic health centers is the development of a centralized electronic repository based on patient data from hospitals in which laboratory measurements, vital signs, longitudinal measurements and claims data are collected. The availability of such data warehouses should theoretically give investigators the ability to conduct in silico observational studies. With their increasing availability, there will be increasing reliance on causal inference methodologies for evaluation of various treatments. In this talk, we discuss three issues in this setting. The first is an inherent tension between covariate overlap, treatment positivity and regular estimation procedures with high-dimensional confounders. The second is the use of Bregman distances for achieving covariate balance for improved causal effect estimation. Finally, we consider the role of database search techniques for the use of causal inference in the big data setting.