The first session on this new journey will be a talk by David MacGillivray about Hierarchical Models.
David, who has been presenting many of the topics on our reading journeys last year, as well as his own data modeling projects, will teach the topic through an example from the paper Bayesian hierarchical model for the prediction of football results by Gianluca Baio and Marta A. Blangiardo. ⚽
Reimplementing and exploring some of the paper’s methods using PyMC (version 5), David will demonstrate some of the joys, challenges, and practices of Hierarchical models. We will see a little bit of what might go wrong, as well as some common solutions.
Time This session will be repeated twice, to welcome different time zones:
Wednesday, August 16th, 2pm UTC
Saturday, August 26th, 4:30pm UTC
As usual, the lessons of the first session will probably result in further exploration and polish before the second one.
Length The sessions will be 90 minutes long. Some of us may wish to stay afterwards and chat.
Recording Some parts of the sessions are recorded and shared internally in the Zulip chat. Possibly, we will also share one of the sessions publicly.
Assumed Background For this session, we will assume participants have familiarity with probabilistic programming in PyMC. We will also assume familiarity with core ideas of Bayesian Statistics, equivalent to Chapters 1,2,3,4 of the Bayesian Computation book.
Joining If you wish to be added to our calendar events, please refer to the Joining Jointprob form.
- Data Preparation
- Model 1 - Hierarchical - direct translation from original paper
- About League Football
- Forward Model
- Model 2 - Unpooled Model
- Why Hierarchical?
- Model 3 - Revisit original hierarchical model, but with modified prior distributions
- Model 4 - Non-centred Hierarchical Model
- Analysis of Model 4 Results
- Shrinkage
- Summary
- Resources