Hierarchical Bayesian Modeling of Survey Data with Post-stratification

December 08, 2022
By Thomas Wiecki
Introduction
In this panel discussion, Tarmo Jüristo tells us how Bayesian modeling can help in environments where data are noisy and uncertainty is high –like public opinion polls. In particular, data can be sparse in some strata of the population, making the model’s job harder, precisely for the demographics you’re the most interested in. A special focus is placed on the work PyMC Labs has done with Tarmo, implementing a state-of-the-art hierarchical Bayesian model. Coupled with post-stratification, this method not only makes inference possible – it makes it actionable, even you have only a few data points for some demographics.
Timestamps
00:00 Introduction by Thomas
03:45 Tarmo introduces himself
05:20 Panel discussion starts
06:11 Description of Salk
08:13 Zooming into the data Salk uses
10:04 A look into what Tarmo does
13:58 Multilevel regression with post-stratification
16:27 Further refinements of the Multilevel regression with post-stratification
19:57 Model output
25:50 Question: On a multilevel aspect, does this mean you model other clusters/groups within other clusters/groups?
28:43 Input to simulation
32:20 Final simulation
34:46 Alex Andorra introduces himself
36:40 Question: How do you choose whether it makes sense to add interactions to a model and do you start with all possible interactions?
38:56 Technical difficulties during the project
46:59 Demonstration of the dashboard
51:52 You can use geospatial covariation to extend the model
53:27 Does the forecasting take the difference in policies between parties
54:19 Using Gaussian Processes in the model(Advantages and disadvantages)
59:55 Question: If you have more time, what would you add to the model
1:02:56 Question: How well do you think the model is taking without rare events?
1:06:57 Thank you!