Valentin Zulj: Contributions to Statistical Model Averaging

On the 10th of October, at 09:15 in Lecture Hall 2, the thesis defence of Valentin Zulj will take place, with the thesis "Contributions to Statistical Model Averaging".

This thesis studies and develops theory and methodology concerned with compromise estimation in prediction problems.

Paper I considers frequentist model averaging of predictions made using generalized linear models. The paper includes proofs of asymptotic properties concerning the predictive performance of the averages, as well as a simulation study where averages are evaluated. A special focus of the paper is models estimated by penalized maximum likelihood.

Paper II works in the context of causal inference, and studies the use of model averaging to combine candidate estimates of the propensity score. The paper proposes a weighting criterion designed to simultaneously target the accuracy and balancing properties of the combined propensity score, and evaluates the criterion in a simulation study.

Paper III considers Bayesian stacking of linear and logistic regression models, and derives frequentist asymptotic properties of averages compromising between Bayesian regression models. The paper also applies Baysian stacking in simulation.

Paper IV considers the estimation of frequentist model averages from an alternative perspective. Here, the focus lies on estimating candidate models conditioned on a given set of fixed weights, rather than estimating weights conditioned on a given set of candidate models. Two procedures are outlined, and applied in simulation.