Data-adaptive learning in causal inference for high-dimensional data structures

点击翻译按钮获取中文摘要

Causal inference is the study of the effects of interventions on an outcome. For instance, we might be interested in evaluating the consequences of a carbon reduction environmental initiative or comparing two different treatment approaches for managing an illness. Mathematically, differences in effects are easy to compare when we perform a randomized experiment. Randomized experiments involve the recruitment of study participants/units who are then randomly allocated one of two interventions. Since the allocation is random, we can compare the results for each intervention group at the end of the study, and any difference between the two groups should be due to the difference in intervention. In particular, due to randomization, the two groups of subjects should look very similar in their characteristics before the intervention is applied. However, in the absence of a randomized experiment, subjects are not randomly assigned to intervention groups, but are differentially exposed to interventions, potentially based on their predispositions. This type of data (called observational since we are not experimentally intervening) is much more complicated to analyze. ***My research program focuses on the effects of interventions when the only data collected is observational. In this circumstance, I am particularly interested in developing estimators for the effect of interventions that vary over time. One of the complications in evaluating such data is that many factors can influence which subjects experience an intervention at a given time-point, and these factors also affect the study outcome. Special models called "Marginal Structural Models" were developed to represent the effects of interventions in the scenario where the interventions were randomly allocated. We can estimate these models even when data is observational by adjusting for the influential factors. My research program involves developing estimators of these models when there are a very large number of factors that need to be adjusted for, and when interventions can change over time. The estimators that I want to develop will estimate effects of interest very efficiently (i.e. they will better target what we want to estimate than competing estimators) and are robust to mistakes that we make when assuming the structure of certain models (i.e. they will still do a good job even if some of our assumptions are incorrect).**