Course Schedule and Content 2024
Schedule Overview
The course is taught in period V (22 April – 3 June 2024) on Mondays from 14:15-16:00 in R030/A133 T5 in CS building.
| Seminar Slides | Date | |
|---|---|---|
| Session 1 | Slides | 22 April 2024 |
| Session 2 | Slides | 29 April 2024 |
| Session 3 | Slides | 06 May 2024 |
| Session 4 | Slides Example Example Source | 13 May 2024 |
| Session 5 | Slides | 20 May 2024 |
| Session 6 | Slides | 27 May 2024 |
| Session 7 | Final presentations | 3 June 2024 |
Schedule Detail
Session 1: Introduction to Bayesian Workflows
Learning Outcomes for the Session
- Understanding of why data analysis / statistical / Bayesian Workflows are needed (i.e., current problems)
- Basic understanding of how Bayesian Workflows aim to solve problems
- Basic understanding of steps in workflow
What the Session will Cover
- Practicalities of course
- Details of the grading/assessment
- Timing/schedules
- Where to find more information and support
- Relationship to prior BDA project tasks
- Concepts/tasks from BDA which will be relevant
- Possible datasets and modelling problems
- For students without their own
- Exploratory data analysis in Bayesian Workflow
Recommended Resources for Next Session
- A preprint on “Bayesian Workflow” by Gelman et al. (2020) arXiv
- A talk on iterative model building and Bayesian workflows by Aki Vehtari: video recording
- Section 1-3 and 4.1 in Aki’s case study “Birthdays workflow example” (https://users.aalto.fi/~ave/casestudies/Birthdays/birthdays.html)
- A talk on “An introduction to Bayesian multilevel modeling with
brms” by Paul-Christian Bürkner at Generable: video recording
Session 2: Choosing an Initial Model
Learning Outcomes for the Session
- How to specify a research question that can be answered with a statistical model
- Awareness of tools and methods to aid in exploring data and formulating question
- Awareness of common/standard modelling approaches for different questions
- Students should have decided on their research question and at least 1 initial model
What the Session will Cover
- Using exploratory data analysis to support the choice of initial models
- Common models and modelling approaches for common research questions
- e.g., observational, randomised study, purely exploratory, based on theory
- Literature and best-practices can also help with model development
- Using either an example or a student’s dataset/problem as an example
Recommended Resources for Next Session
- BRMS demo for generating from prior
- Stan prior choice recommendations
- Stan distribution visualiser
- Distribution explorer
- PreliZ
bayesplotpackage- A talk on “Why not to be afraid of priors (too much)” by Paul-Christian Bürkner at Bayes@Lund 2018: slides video recording
Session 3: Prior Choices
Learning Outcomes for the Session
- Develop awareness of impact of mis-specified priors
- Understanding of approaches for specifying prior (i.e., prior elicitation)
- Understanding of tools/methods for diagnosing prior sensitivity
What the Session will Cover
- How to turn assumptions/knowledge into prior
- Generative priors, penalised complexity, etc.
- Connection to model expansion/choice/selection goals
- Brief coverage of different topics for prior choice
- How to assess prior choice
- If the type of prior / prior properties do not align, etc.
- Prior predictive checks
- How to revise/modify a prior (if needed)
Recommended Resources for Next Session
Session 4: Model Checking: Posterior Predictive Checks & Calibration
Learning Outcomes for the Session
- Understanding role of predictive checks in model-checking
- Understanding of impact of different data types on approach (e.g., continuous vs discrete)
- Familiarity with different graphical methods and tools for supporting checking
What the Session will Cover
- Detecting, resolving, and reporting:
- Prior sensitivity
- Posterior predictive checks
- Calibration
Recommended Resources for Next Session
Session 5: Extending Models and Model Selection
Learning Outcomes for the Session
- Understanding of how to extend a model to better address a research question, and if it is even needed
- Understanding of how to select between different models, and whether this is necessary
- Understanding of how to combine multiple models for increased performance
What the Session will Cover
- Common methods for model expansion
- Data-driven vs theory-driven
- Common methods for model comparison, and interpreting the results of these
- Alternatives to model selection
- e.g., model averaging
Recommended Resources for Next Session
- Poststratification in Stan User’s Guide
- A talk on “A biased tour of the uncertainty visualization zoo” by Matthew Kay: video recording
- Case study by Aki Vehtari on reporting the correct number of digits
- A paper on “The only thing that can stop bad causal inference is good causal inference” on by Rohrer, Schmukle, & McElreath (2022) link to pdf
Session 6: Interpreting and Presenting Model Results
Learning Outcomes for the Session
- Using model quantities and results to assess whether research question has been answered
- Reflecting on how/if initial question has changed throughout workflow process
- Reflection on possible alternative expansions for research questions/models
- Presenting model results accessibly
What the Session will Cover
- How to extract and prepare results
- e.g., using R packages such as
tidybayesandmarginaleffects - Further analysis of interesting effects (interactions, smooths, etc.)
- e.g., using R packages such as
- Prior sensitivity for final conclusions (quantities/choices not sensitive to priors)
- Brief introduction to alternative methods for constructing and reporting alternative models
- e.g., Multiverse analysis
Session 7: Summary & Presentations
Learning Outcomes for the Session
- Awareness of how workflow will differ between research questions
- Reflection of how workflow contributed to analysis process
- Reflection on future iterations of model