Building Models

Routing Summary

Part 2 (Chapters 5–7) of Gelman, Vehtari & McElreath (2026): choosing and specifying a model, every prior-choice rule the book states, simulation as a modeling tool, and the extension from inference to generalization, causal inference, and decision. 19 notes.

Concept Map

ConceptNoteTypeDepends OnKey Result
Where to start a modelChoosing an Initial ModelconceptFour Modeling ScenariosStart simple enough to fit, complex enough to be relevant
Assumptions → modelRelating a Model to Subject-Matter AssumptionsconceptChoosing an Initial ModelModel structure should encode substantive claims
Formal specificationExpressing a Bayesian Model with Probability Distributionsdefinition—Joint and its decomposition
Data model ≠ likelihoodA Data Model Is Not Just a LikelihoodconceptExpressing a Bayesian Model with Probability DistributionsThe same plays two roles
Generative vs. partially generativeGenerative and Partially Generative ModelsdefinitionA Data Model Is Not Just a LikelihoodOnly fully generative models support prior predictive simulation
Prior choicePrior DistributionsconceptExpressing a Bayesian Model with Probability DistributionsWeakly informative as default; the flat prior is not neutral
Effect-size scalingConstructing Priors for Effect SizesconceptPrior DistributionsStandardize, then place the prior on an interpretable scale
Joint priorsJoint Priors and Covariance MatricesdefinitionPrior DistributionsLKJ for correlation matrices; horseshoe and R2D2 for coefficient vectors
Data model + prior togetherSpecifying the Data Model and the PriorconceptPrior DistributionsThe pair, not either alone, defines the model
Modeled vs. unmodeled dataModeled and Unmodeled DatadefinitionExpressing a Bayesian Model with Probability DistributionsWhat you condition on vs. what you model
Prior predictive checkingPrior Predictive CheckingconceptGenerative and Partially Generative ModelsSimulate from the prior; look at the implied data
Prior-likelihood conflictTail Behavior and Prior-Likelihood ConflicttheoremPrior Predictive CheckingTail behavior determines which source wins under conflict
Simulation for uncertaintySimulation to Express Uncertaintyconcept—Propagate draws rather than combining standard errors
Point estimatesPoint Estimates and UncertaintiesconceptSimulation to Express UncertaintyWhen a summary is enough and when it misleads
Simulated-data experimentsDesigning Simulated-Data ExperimentsconceptSimulation to Express UncertaintyDesign the experiment, not just the simulation
Full-pipeline simulationSimulating an Underlying Process, Data Collection, and InferenceexampleDesigning Simulated-Data ExperimentsSimulate process, collection, and inference together
PoststratificationPoststratificationdefinitionSimulation to Express UncertaintyMRP: model at the cell level, reweight to the population
Causal inferenceCausal Inference as GeneralizationconceptPoststratificationSATE/PATE as prediction under counterfactual predictors
DecisionFrom Inference to DecisionconceptCausal Inference as GeneralizationExpected utility averaged over the posterior

Notes

Sources

  • Gelman Vehtari McElreath 2026 - Bayesian Workflow (book) — Chapters 5–7, pp. 63–134

See Also