Flexible estimation of causal effects with observational data

Details

Supervisors

Pircalabelu, Eugen

Faculty

Faculté des sciences

Degree label

Master [120] en statistique, orientation générale, à finalité spécialisée

Abstract

In many situations, researchers want to evaluate the causal effect of an intervention on an outcome of interest. The gold-standard to address such questions is to run a randomized controlled trial. However, running a randomized trial is not always possible, and researchers sometimes must rely on observational data to attempt to answer these questions. When relying on observational data, identification of causal effects is subject to an identification problem due to the presence of confounders. The identification problem can be solved by making additional identifying assumptions. When it comes to estimation, the relyance on these additional assumptions often compel researchers to use more sophisticated statistical methods. Popular approaches for doing so rely on the specification of a model for the process whereby the data were generated. However, the final result can be sensitive to the modelling choices. This has encouraged researchers to develop more flexible methods to evaluate causal effects that relax these modelling assumptions. Some of these methods involve the use of very flexible data-adaptive estimation procedures. However an important problem when using these more flexible methods relates to the understanding of the statistical properties of the estimators. Recent advances have been made in understanding the properties of estimators for causal quantities that involve flexible data-adaptive methods. In this thesis, we review and discuss these recent advancements in comparison with the more traditional approaches to the estimation of causal effects.