Learning Bayesian Statistics BITESIZE | Why Your Models Might Be Wrong & How to Fix it, with Sean Pinkney & Adrian Seyboldt
Jun 4, 2025
This discussion features Sean Pinkney, an expert in statistical modeling, alongside Adrian Seyboldt. They explore the concept of Zero-Sum Normal in hierarchical models and its implications. The duo dives into the challenges of incorporating new data, distinguishing between population and sample effects, and offers insights into enhancing model accuracy. They also suggest potential automated tools for improved predictions based on population parameters, tackling common statistical modeling challenges along the way.
AI Snips
Chapters
Transcript
Episode notes
Zero-Sum Normal Improves Sampling
- The zero-sum normal constraint simplifies sampling and often yields better estimates in hierarchical models.
- It allows better inference of population mean effects and covariances despite parameter constraints.
Use Population Parameters for New Groups
- Use learned population parameters to predict new group members even if they weren't observed in fitting.
- Incorporate informed priors for new group elements to improve predictions.
Distinguishing Sample vs Population Effects
- Effects can be distinguished relative to the sample mean or the population mean.
- Zero-sum effects inherently center around the sample mean, impacting interpretation in finite populations.