Paul Messer

// DISSERTATION PROJECT

Machine learning within SAE: Gradient Boosting with random intercepts for flexible domain prediction

Estimates of spatially disaggregated indicators at small area-specific sample sizes, so-called small area estimates, are often based on regression approaches of linear mixed models. In contrast to this, machine learning methods offer nonlinear and nonparametric alternatives that combine promising predictive performance and lower risk of model misspecification.

Among these methods are the mixed-effects gradient boosting methods presented in this paper, which combine the advantages of gradient boosting decision trees with the ability to model hierarchical dependencies. These methods are used in this case to estimate small area means.

The proposed methods are evaluated in model-based simulations and on the basis of design-based simulations of Mexican income data from the state of Nuevo León.

First results show promising improvements compared to typical small area regression-based approaches (EBP, BHF).

........................................................................................................................................................

// ACADEMIC BACKGROUND  

2021 – ongoing      
PhD Candidate in Statistics at Otto-Friedrich University in Bamberg

2019 – 2021
Master of Science in Survey Statistics at Otto-Friedrich University in Bamberg

2016 – 2019
Bachelor of Science in Statistics at LMU in Munich

........................................................................................................................................................

// EXPERIENCE

Lecturer at the Chair of Statistics in Bamberg (Advanced Data Analysis with R, Introduction to Programming with R, Introduction to Statistical Programming with Python, Statistical Analysis of Missing Data)

Statistical Consultant at BACES

Student Employee, Junior Data Scientist at quantified markets GmbH