Assumption Glossary¶
Every causalrl Certificate records the assumptions it consumed as typed Assumption entries
(name, params, checkable, diagnostic). This page explains each one: what it means, whether
it is checkable from data, and how a violation is handled. The library is one-sidedly honest —
outside a supported class it returns a hedge or a typed exception, never a silent point estimate.
Epistemic kinds¶
A certificate's kind is never conflated across these three:
IDENTIFIED— a point claim backed by identification under the stated graph/assumptions.BOUNDED— partial identification: an interval valid under an explicit sensitivity budget.EMPIRICAL— simulation/sample evidence only, with no identification guarantee.
Assumptions¶
backdoor¶
The adjustment set blocks all back-door paths from treatment to outcome. params.adjustment_set
lists the covariates. Checkable given the graph (the graph is the witness); the numeric estimate
additionally assumes correct nuisance models (relaxed by the doubly-robust aipw / dml methods).
overlap (positivity)¶
Every unit has propensity bounded away from 0 and 1 (or, for importance sampling, a finite effective
sample size). params.eps / params.min_ess_fraction set the threshold; diagnostic reports the
observed extremes / ESS. Checkable — a violation downgrades to a hedge (overlap-violation)
rather than returning an unstable inverse-weighted estimate.
MSM (marginal sensitivity model)¶
Unmeasured confounding shifts the true propensity from the nominal one by at most an odds-ratio
gamma >= 1 (Tan's model). params.gamma is the budget. Not checkable from data — it is a stated
sensitivity parameter; the certificate is BOUNDED and the interval widens monotonically in
gamma, collapsing to the point estimate at gamma = 1.
logged-propensities¶
The importance weights / nominal propensities used for off-policy evaluation are the true logging
probabilities (up to the MSM budget when one is declared). Not checkable from the logs alone.
moment-condition¶
The target functional's moments exist (e.g. a finite mean). Checkable via a heavy-tail diagnostic;
when it fails, the mean target downgrades to a valid quantile/tail target with the downgrade
recorded (hedge.downgraded_from).
mi-cap / pivotality¶
A cap on how much a hidden confounder can move the decision, expressed as a mutual-information or
odds-ratio budget (the pivotality layer). BOUNDED; certify_decision reports the tipping gamma
at which the decision would flip.
selection-nodes-S (transport)¶
The named selection nodes mark exactly where the source and target regimes differ (population or
mechanism). The transport formula / regret certificate is valid when the marked diagram is correct;
Regime selection nodes are the witness. Non-transportable queries hedge.
quantile-sketch¶
A streamed quantile is answered by a Greenwald–Khanna sketch whose true rank is within epsilon * n
of the requested rank. params.epsilon is the guaranteed rank-error budget, recorded in provenance
so the approximation travels with the claim.
fqe-model¶
A fitted-Q-evaluation off-policy value estimate. Model-based with no identification guarantee under
hidden confounding, so the certificate is EMPIRICAL; upgrade to certify_effect or bound it with
certify_policy when the structure licenses it.
Provenance¶
Every certificate also records reproducibility metadata (Provenance): the library version, seeds,
a data fingerprint, a graph hash, and a timestamp — so a claim can be reproduced and audited.