Skip to content

External prior bank

Transfer previously fit posteriors into a new analysis as power / mixture priors — without pretending priors create nonparametric identification.

Surveys are the usual motivating domain (product / context tags), but the library speaks in sources, targets, and designs.

Workflow

catalog.filter(target) → rank(similarity)
  → map(effect|named params) → power-prior / mixture (α_k, w_k)
  → apply transport policy → PriorSet + assumptions
  → analyze(..., inference=Bayesian(prior_from=…))
  → prior PPC + prior sensitivity (α grid when banked)
  1. Wrap posterior artifacts with PriorSourceMeta / PriorSource.
  2. PriorCatalog.compatible_with(...) → accept / partial / reject reasons.
  3. Callers supply similarity scores; the library ranks but does not invent domain similarity.
  4. compose_external_priors(...) builds a ComposedPrior (optional ConflictPolicy, TransportPolicy).
  5. Pass the composed prior as Bayesian(prior_from=...).
  6. With refute="full", the ATE path attaches an α-multiplier sensitivity grid (not isotropic scales) when an external compose is present.

Python example: python/examples/prior_bank_surveys.py.

Invariants

  • Priors are recorded as PriorRestriction assumptions; they never upgrade IdentificationStatus.
  • Heterogeneous designs transfer at the effect-functional (or explicitly mapped parameter) level by default — not silent coef_i → coef_i. When Bayesian(prior_from=artifact) leaves mapping unset, hydrate chooses identical subspace for matching designs and EffectFunctional when layouts differ and an effect quantity exists; otherwise an explicit mapping is required.
  • Unmapped / incompatible mass goes to a weakly informative baseline; never silently renormalized.
  • Conflict can only shrink external weight (α → 0), never invent identification or increase α.
  • Population / environment shifts require an explicit TransportPolicy; missing policy → structured transport_policy_required. Pass prior_sources= (or source_populations=) with target_population= so compose reads tags["population"] via populations_from_prior_sources — callers need not thread population tags manually when catalog meta is available.

What callers must supply

Concern Owner
Similarity / ranking scores Caller
Product / context / taxonomy tags Caller conventions on tags
Max-trust / governance of the bank Caller
Compatibility, mapping, α/weights, transport assumptions, diagnostics Library

Exact tag keys on compatible_with(tags=...) are hard filters. Soft similarity belongs in catalog.rank(scores=...).

Supported transport policies

Documented in parity bayes.prior_bank.transport:

  • InvariantConditionalOutcome — P(Y|do(T),X) stable across populations
  • InvariantEffectModifiers
  • InvariantPropensity (without transport weights → α forced to 0)

Unsupported environment / unidentified shifts are not claimed.

Gates

bash scripts/gate_bayesian.sh

Runs conformance/bayesian/prior_bank_* plus python/tests/test_prior_bank.py.