Prior Elicitation for PyMC Models

Decision Flowchart: Choosing a Prior Strategy

Do you have domain expertise or expert access?
├── YES: Can the expert quantify beliefs precisely?
│   ├── YES → Expert elicitation (SHELF protocol, PreliZ roulette/quartile)
│   └── NO → Constrained priors (find_constrained_prior, PreliZ maxent)
└── NO: Do you know the plausible scale of the parameter?
    ├── YES → Weakly informative priors (Normal, HalfNormal, Student-t)
    └── NO → Use prior predictive checks to calibrate
        └── Generate predictions → Do they cover plausible outcomes?
            ├── YES → Prior is acceptable
            └── NO → Tighten or widen prior, repeat

When to Use Each Strategy