A Mermaid Diagram of Bayesian Workflow

In this post I visualize the Bayesian workflow introduced by Gelman et al. (2020).

I am setting Quato’s echo: true so that you can see the Mermaid instructions for the diagram.

flowchart TD
1["Pick an initial model"]
2{"Prior predictive check"}
3["Fit the model"]
4{"Validate computation"}
5["Address computational issues"]
6["Evaluate and use model"]
7["Modify the Model"]
8["Compare Models"]

1 --> 2
2 --> |Accept Priors| 3
2 --> |Reject Priors| 7
3 --> 4
4 --> |Invalid| 5
4 --> |Valid| 6
5 --> 3
6 --> |Reject Model| 7
6 --> |Accept Model| 8
6 --> |Accept Model| 7
5 --> |Give Up| 7
7 --> 2

flowchart TD
1["Pick an initial model"]
2{"Prior predictive check"}
3["Fit the model"]
4{"Validate computation"}
5["Address computational issues"]
6["Evaluate and use model"]
7["Modify the Model"]
8["Compare Models"]

1 --> 2
2 --> |Accept Priors| 3
2 --> |Reject Priors| 7
3 --> 4
4 --> |Invalid| 5
4 --> |Valid| 6
5 --> 3
6 --> |Reject Model| 7
6 --> |Accept Model| 8
6 --> |Accept Model| 7
5 --> |Give Up| 7
7 --> 2

In the previous diagram I left out details of what to consider in some of the steps. The following diagram expands on some of the steps by showing there are a variety of different approaches depending on the current state of a workflow.

mindmap
  root((Bayesian Workflow))
    Validate Computation
      Convergence Diagnostics
      Fake data simulation
      Simulation based calibration
    Evaluate and use model
      Posterior predictive check
      Cross validation
      Influence of individual data points
      Influence of prior
      Prediction
      Post stratification
    Modify the model
      Pick a new starting model
      Replace model component
      Enrich/Expand model
      Use an approximation
      Add more data
      Modify priors
    Compare models
      Comparing inferences
      Multiverse analysis
      Model averaging/stacking
    Addressing computational issues
      Simplify the model
      Implement model components separately
      Run small number of iterations
      Run on a subset of the data
      Stacking individual chains
      Check for multimodality
      Reparametrize
      Plot intermediate quantities
      Add prior information
      Add more data
      Give up

mindmap
  root((Bayesian Workflow))
    Validate Computation
      Convergence Diagnostics
      Fake data simulation
      Simulation based calibration
    Evaluate and use model
      Posterior predictive check
      Cross validation
      Influence of individual data points
      Influence of prior
      Prediction
      Post stratification
    Modify the model
      Pick a new starting model
      Replace model component
      Enrich/Expand model
      Use an approximation
      Add more data
      Modify priors
    Compare models
      Comparing inferences
      Multiverse analysis
      Model averaging/stacking
    Addressing computational issues
      Simplify the model
      Implement model components separately
      Run small number of iterations
      Run on a subset of the data
      Stacking individual chains
      Check for multimodality
      Reparametrize
      Plot intermediate quantities
      Add prior information
      Add more data
      Give up

Happy modelling.

References

Gelman, Andrew, Aki Vehtari, Daniel Simpson, Charles C. Margossian, Bob Carpenter, Yuling Yao, Lauren Kennedy, Jonah Gabry, Paul-Christian Bürkner, and Martin Modrák. 2020. “Bayesian Workflow.” https://arxiv.org/abs/2011.01808.