sunxd3 Fake-Data Simulation
Pick known parameters, simulate data from the model, refit, and check that the inference recovers what you put in.
Variants
references/single-draw.md— one set of true parameter values, simulate once, refit, check that the posterior covers the truth. Cheap. Catches obvious model bugs and Stan-program errors.references/sbc.md— Simulation-Based Calibration (Talts, Betancourt, Simpson, Vehtari, Gelman, 2018). Many draws from the prior, refit each, check rank uniformity. Rigorous. Validates the inference algorithm's calibration, not just whether one truth can be recovered.references/decision.md— PASS / FAIL criteria and common failure modes (model misspecification, non-identifiability, computational geometry).
Default to single-draw for the pre-fit gate in the workflow. Escalate to SBC when single-draw passes but you suspect inference miscalibration, or when the modeling choice is high-stakes enough to warrant rank-based validation.
Key Practice — Stan is the single source of truth
All data generation must happen in Stan. Python chooses parameter values, passes them to Stan as data, and reads back the simulated observations. Python must NOT implement the generative model — no numpy.random, no scipy.stats for generating synthetic observations. If the simulator is written in Python, fake-data simulation validates Python-to-Python consistency, not the Stan model itself.
The simulator.stan file is a generated-quantities-only Stan program (no parameters block, no model block). Its transformed-parameter computation and _rng calls must be a line-by-line mirror of the corresponding blocks in model.stan. See references/single-draw.md for the canonical pattern.