Latent Variable Models in the Deconfounder; a Multiple Causal Inference Algorithm
Specialeforsvar: Anne Sophie Roed
Titel: Latent Variable Models in the Deconfounder; a Multiple Causal Inference Algorithm
Abstract: Causal discovery on observational data, especially when prediction falls short, is an important problem and is to this day an active field of research. Causality problems on real-world data is most often threatened by unobserved confounders and this problem is increasingly problematic as the number of causes expands. The research often focus on a single cause and outcome. This is on the expense of doing multiple causal inference, which would be preferred on high-dimensional real-world data. Recent research challenges this sometimes limited sphere of causal discovery on observational data and argues in favour of latent variable models as a model class that can estimate good substitute confounders for the unobserved multiple cause confounders. This brings multiple causal problems on a par with single causal problems on observational data with potential unobserved confounders. The pivoting point of this thesis is the use of latent variable models in the Deconfounder Algorithm presented by Wang and Blei (2019). We introduce a general and consistent framework for latent variable models and methods. This generalization allows us to build, compute, and check flexible models that can successfully quantify and estimate the hidden structures in data. We asses the various methods on real-world data and discuss the findings, which points towards that the Deconfounder Algorithm "deconfounds".
Vejleder: Niels Richard Hansen
Censor: Nina Munkholt Jacobsen