Learning Bayesian Statistics #34 Multilevel Regression, Post-stratification & Missing Data, with Lauren Kennedy
Feb 25, 2021
Lauren Kennedy, a Business Analytics lecturer at Monash University, discusses the complexities of multilevel regression and post-stratification (MRP) for analyzing non-representative data. She shares insights on how structured priors can enhance demographic analysis, addresses the challenges of missing data imputation, and highlights the importance of causal inference in social sciences. Additionally, Kennedy emphasizes teaching Bayesian methods through practical workflows, ethical considerations in data analytics, and the necessity for inclusivity in statistical research.
AI Snips
Chapters
Books
Transcript
Episode notes
Real-World Survey Non-Response
- Lauren Kennedy shares real examples of non-response bias in survey data, like missing gender responses due to limited options.
- She highlights how surveys and social science data uniquely suffer from non-representativeness challenges.
MRP Model Overcomes Simple Weighting
- Multilevel Regression and Post-stratification (MRP) models relationships between outcome and demographics rather than reweighting samples directly.
- MRP pools information via multilevel models, regularizing estimates and improving prediction for population-level inference.
Prior Predictive Checks Essential
- Always perform prior predictive checks to understand your priors and their impact before fitting Bayesian models.
- Avoid using priors from other studies blindly; tailor them to your context and link function.
