axiom

Reference

Glossary

Short definitions, in the vocabulary this library uses: treatment, dose, unit, outcome, covariate. Where a term has a guide of its own, the definition here is the one-line version and the link is the real answer.

Causal structure

What causes what

causal graph (DAG) A set of variables and arrows, where an arrow means “intervening here could change that”. The missing arrows carry the content. Guide →
confounder A common cause of treatment and outcome. Creates a correlation that is not an effect. Must be adjusted for. Guide → adjust
collider A common effect of two variables. Adjusting for it creates a correlation that was not there — the reason “control for everything” is bad advice. never
mediator A variable on the causal path from treatment to outcome. Adjusting for it leaves the direct effect only — a different question. depends
instrument A variable that moves the treatment and reaches the outcome by no other route. A lever, used when the back door cannot be closed.
d-separation The graphical rule for reading conditional independence off a DAG. What every identification criterion on this site is ultimately checking.
do(X = x) Setting a variable from outside, breaking whatever normally determines it. Different from observing that it equals x, and the whole difference between prediction and intervention. Guide →

Identification

Whether the effect is recoverable at all

identification Whether a causal quantity can be recovered from the available data at all, given the graph. Prior to, and independent of, estimation. Guide →
back-door criterion Block every non-causal path by conditioning on a set containing no descendant of the treatment. The most common route.
front-door criterion Recover the effect through a fully mediating variable, even when the confounder is unmeasured.
adjustment set The variables to condition on. Derived from the graph, not chosen by intuition. Several valid sets are a consistency check, not a menu.
status: identified / downgraded / blocked A licensed route with everything measured; a route whose requirements are unmeasured or unverified; no route in this graph. Most real work is downgraded, correctly.
assumption In axiom, a named object with a statement, a state (satisfied / unverified) and the thing that would refute it. Not a footnote.
sensitivity analysis How strong an unmeasured confounder would have to be to overturn the conclusion. Report it with the estimate. Guide →

Defining the quantity

What exactly is being estimated

estimand The quantity you want, written down completely: eight facets, all required. Distinct from the estimator (the procedure) and the estimate (the number). Guide → 8 facets
facet One of the eight things an estimand fixes: quantity, intervention, outcome, population, window, level, conditioning, dimension. Two quantities are the same only if all eight match.
contrast / marginal / elasticity Different functionals of the same response curve — a difference between two doses, a slope at a dose, a percentage response. Not interchangeable, and axiom refuses to transfer between them.
dimension / unit / scope What kind of thing (currency, time, outcome); which measure of it (USD, days); and which population, window and level. Three separate mismatches. Guide →
level The unit of analysis — individual, cluster, aggregate — together with the interference model. A household is not a person.
interference When one unit's treatment affects another unit's outcome. If it is present and undeclared, most standard analysis is wrong; axiom blocks a level transfer that would assume it away.

Carrying an answer somewhere else

Transportability

transportability Whether an effect measured in one population is valid in another. A different question from identification, with different criteria. Guide →
selection diagram A causal graph plus an S[·] node on every variable whose mechanism differs between the two populations. Turns “different population” from an adjective into a testable structure.
S-admissible set A set Z that separates the S-nodes from the outcome, licensing the transport formula. Here: age.
transport formula The source population's Z-specific effects, reweighted by the target's own distribution of Z: P*(illness|do(filters)) = Σ_{age} P(illness|do(filters),age) P*(age)
trivial transportability The target population identifies the effect from its own data, so the source study is not needed at all. Good news that reads like a technicality.
transfer plan What Estimand.transfer_to returns: every differing facet with the assumption that would license it, or a refusal. No facet passes silently.
unsupported None of the implemented sufficient conditions fires. A claim about the evidence, deliberately not a claim that no transfer exists. not “no”

Keeping the trail

Provenance

spec Any declarative configuration object in axiom: frozen, versioned, content-hashed, and serializable to JSON and back. Guide →
content hash A stable address for a spec's contents, independent of process, machine and dict ordering. Same question, same hash; one changed field, different hash.
ledger / ledger line The record of every boundary a number crossed. One line per differing facet, complete by construction, with facts and assumptions distinguishable at a glance.
schema version Every spec class carries one. Loading an older record applies a registered migration or raises naming both versions — it never loads wrong.

Design, fitting and evidence

Doing the work

prior A claim about magnitudes in the outcome's own units, statable to a domain expert. Not a regularization setting. Guide →
prior predictive check Simulate from the prior, push it through the model, and look at the outcomes it implies. Catches most bad priors before any data.
MDE Minimum detectable effect. Conventionally the input to sizing; better treated as an output, derived from the boundary the decision turns on. Guide →
EVPI / EVSI Expected value of perfect information, and of this specific sample. What the study is worth in the decision's own currency — the ceiling, and the actual.
calibration Folding an experimental measurement into a broader model, as a prior, a likelihood term, or a constraint. The route changes the answer. Guide →
carryover How a dose's effect spills into later periods. Decides how much of an effect a short observation window can see, and therefore the window facet's correction.
response surface The fitted mapping from dose to outcome. Everything — the likelihood, the design maths, the optimizer — reads it through one forward().
design effect How much a clustered sample inflates a standard error relative to an independent one. Why reading a household measurement per person widens the interval.
meta-analysis Pooling evidence across studies with their heterogeneity taken seriously. The pooled interval is usually wider than any single study's.
stopping rule A pre-committed schedule of looks and the boundary each must cross. Without one, “stop when it's clear” has no error rate.

Looking for a specific function or class instead? The API map lists every public symbol in the library with its signature and a real call taken from an executed notebook.