Multilevel regression and post-stratification (MRP), the current gold standard, faces limitations in extrapolating survey data to smaller geographic units. This paper proposes BART-based methods—specifically Bayesian Additive Regression Trees for Prediction (BARP)—which offer significant advantages through advanced nonparametric regularization techniques.
Methodology & Findings:
Through systematic comparison across diverse datasets,
BARP consistently outperforms traditional MRP, demonstrating superior accuracy in estimating opinions at targeted geographical levels. The BART approach provides a more flexible alternative for extrapolating survey results without sacrificing methodological rigor.
Practical Application:
The author provides an accessible R package that makes BARP available for practical use.
This represents a substantial enhancement to survey data analysis methodologies.
| BARP: Improving Mister P With Bayesian Additive Regression Trees was authored by James Bisbee. It was published by Cambridge in APSR in 2019. |
