A Comparison of Deep Hierarchical Models and Machine Learning for MRP

• In Event: Measuring Public Opinion

Abstract

Over the past years, a proliferation of Multilevel Regression with Poststratification (MrP) models has emerged in the literature. These innovative model variants often amalgamate elements from the realm of statistical learning with the foundational principles of classic MrP, leading to advancements that consistently outperform the established benchmarks. A recent contribution (Goplerud 2023) advocates for the incorporation of deep interactions within hierarchical models. It contends that such models can frequently rival the performance of other contemporary variants that are based on machine learning. In this study, we critically evaluate the performance of autoMrP (Broniecki et al. 2021) against deep hierarchical models. Our findings reveal that autoMrP consistently surpasses the performance of deep hierarchical models when operating within a high-dimensional feature space and with a low number of observations. This study not only highlights the strengths of autoMrP in handling complex data structures but also provides insight into the limitations of deep hierarchical models under challenging conditions.

Authors

• Florian Simon Schaffner, University of Zurich

• Philipp Broniecki, University of Oslo

• Lucas Leemann, University of Zurich

• Reto Wueest, University of Bergen