The words, defined rather than assumed.
29 terms covering the causal vocabulary this product uses, the record vocabulary it adds, and the places where its usage is narrower than the literature's.
Most of these words mean something specific in causal inference, and the definition comes first. Where this product uses a term more narrowly than the field does, the entry says so.
Estimate means two things
In the literature it usually means a quantity computed from sampled data, with sampling error. This product does not do that: it computes exact arithmetic over a stated model. The entries are written in the general sense and then say which sense is not available here.
A missing term is a gap
This list is the whole of the vocabulary the site uses. If a word you came for is not in it, that is an omission to fix rather than a term defined on another page — and the search matches definitions, not only headings, so a half-remembered phrase should still find it.
29 of 29 terms
A
- Adjustment setIdentification
The variables a model controls for when computing an effect.
Given as two values: the full backdoor set and the minimal set that satisfies the criterion. The minimal set gives the same estimate with less variance, so a caller adjusting for everything upstream is not wrong — they are paying for variables they did not need. An empty minimal set is a real answer meaning nothing needs adjusting for.
Related: backdoor criterion, collider, mediator
B
- Backdoor criterionIdentification
A test for whether adjusting for a set of variables removes every non-causal path between the treatment and the outcome.
This is the route the engine uses first. It returns the full set that satisfies the criterion and the minimal subset that does, because two callers adjusting for different valid sets get the same estimate with different variance — a real choice that the response makes visible rather than settling silently.
Related: adjustment set, confounder, identification
C
- ColliderIdentification
A variable affected by both the treatment and the outcome.
Conditioning on it can make the treatment and the outcome appear related when they are not, or hide a relationship that is there. It is the most damaging variable to include and the one most likely to be reached for, because downstream outcome measures look exactly like useful controls.
Related: confounder, mediator
- ConfounderIdentification
A variable that influences both the treatment and the outcome, creating a non-causal association between them.
Ancestry alone is not enough, which is a distinction this product got wrong once and now tests for: a variable that reaches the outcome only through the treatment is not a confounder, because there is no backdoor path to block. Its role is reported as a cause of the treatment, and adjusting for it is optional.
Related: adjustment set, collider, mediator
- ContestedProvenanceStructure
An edge the people involved disagree about.
A review flag that says the structure needs a decision, and explicitly not a calibrated probability that the edge is wrong. That is the readiest misreading of the word, and a reader will infer it if the copy leaves room — which is why the definition is written this way.
Related: provenance
D
- Decision recordRecords
What was decided, on what basis, by whom, and what happened afterwards.
Here it is the four things stored separately: the intervention as framed, each approval against the specific stored analysis it was decided on, exposures recorded as assigned and as reached, and the measured outcome reconciled against the design-time claim. Assigned and reached usually diverge, and the divergence is often the most informative number in the record.
Related: intervention, evidence pack
- Discrete Bayesian networkStructure
A set of variables, each with a probability table conditioned on its parents, forming a directed acyclic graph.
The model type. Every variable here is discrete, so a continuous variable reaches the model through binning, which is lossy and labeled on the row rather than hidden. The joint distribution is the product of the tables by construction, which is what makes inference arithmetic rather than simulation.
Related: variable elimination, structure
- do-operatorEstimation
Notation for setting a variable rather than observing it: do(t) means intervene to make t hold, not look at cases where t happens to hold.
The distinction is the whole difference between a causal question and a correlational one. This product computes P(y | do(t)) with the treatment's own incoming edges removed, and states that removal as an assumption on every effect response.
Related: intervention, estimand
E
- E-valueEstimation
How strong an unmeasured confounder would have to be to explain the observed effect away.
Reported as part of the fragility check, with the dependencies it rests on stated: a single binary unmeasured confounder, associated with both treatment and outcome by the same risk ratio, with the effect measure of interest. A large E-value is evidence of robustness against that specific perturbation and not proof that no confounder exists.
Related: refutation, confounder
- ElicitedProvenance
A value obtained by asking — from a document, a person, or a language model reading one.
An elicited value is a claim about how the process works, held by someone. It is not measured, and the tag is what keeps that visible. Note that the tag records where a number came from and not how good it is: a careful workshop and a hurried meeting produce the same one.
Related: fitted, provenance
- EstimandIdentification
The precise quantity a study is trying to measure, stated before any method is chosen.
Most disagreements about an analysis are really disagreements about the estimand. This product fixes it explicitly as a treatment variable, a baseline level, a contrast level, an outcome variable and one outcome state, so the question can be recomputed identically when a check refutes it.
Related: identification, risk difference
- Evidence packRecords
One decision assembled into a single self-contained, fingerprinted document.
The decision, the analysis it was approved against, the exposures and the reconciliation, readable without access to the system that produced it. The fingerprint proves the pack has not been altered since; it says nothing about whether its contents were right, which is the most available misreading of that field.
Related: decision record, replay
F
- FittedProvenance
A value computed from observed counts rather than asserted.
Fitted rows are the only ones with calibrated intervals, which is why the calibrated-only projection is defined over them: the largest sub-network where every row is backed by a fit. A question answered from calibrated tables only is a narrower claim, and a separately available one.
Related: elicited, provenance
- Front-door criterionIdentification
A second route to an effect, which works through a mediator when a confounder was never measured.
It requires a variable that the treatment causes, that causes the outcome, and that nothing unmeasured confounds. The engine checks it whenever the backdoor route fails, and reports the specific reason it does not apply — usually that the treatment has no child leading to the outcome, so there is no mediator to route through.
Related: mediator, identification
H
- HandoffEstimationStructure
Packaging the model for a different tool to estimate on.
The graph as DOT, the adjustment set, and the role of every variable with control advice, in the form an estimating library accepts. The snippet it carries is explicitly labeled as not executed and not verified here, because this service does not depend on those libraries and cannot run them.
Related: adjustment set, collider
- Heterogeneous effectEstimation
An effect that differs across subgroups of the population rather than being one number for everyone.
Computed here by exact elimination per stratum over a discrete model, not learned as a function from rows. The practical difference: no data and no new dependency are required, and the result carries no sampling error — because it is arithmetic on a stated model, which is only as good as the model.
I
- IdentificationIdentification
Whether a causal question can be answered at all from the information available, before anyone tries to answer it.
The question is identified if there is a formula that expresses the effect using only observed quantities. This product reports the verdict and the reason on every effect: identified true or false, the route taken (backdoor, front-door, or none), and a sentence naming why. An unidentified question returns no number, which is a result rather than an error.
Related: backdoor criterion, front door criterion, estimand
- InterventionRecordsEstimation
An action being considered or taken, framed as a decision with options.
This product's record object. An intervention is framed, an option is approved against a specific stored analysis, what was assigned is recorded separately from who was actually reached, and the measured outcome is reconciled against the design-time claim. It is what makes a decision replayable rather than remembered.
Related: evidence pack, decision record
M
- MediatorIdentification
A variable the treatment affects, which in turn affects the outcome — part of how the effect travels.
Adjusting for one reports the direct effect while appearing to report the total, which is the most consequential way to be wrong without noticing. Every variable's role is reported with advice, and a mediator is marked do-not-control.
Related: collider, front door criterion
P
- PositivityIdentification
Whether every combination of variables the analysis relies on actually occurs in the data.
An effect computed in a region where there are no observations is an extrapolation, and it looks identical to one computed where the data is dense. The engine reports which strata an adjustment set has no data for as a separate result rather than folding it into the effect.
Related: adjustment set, identification
- ProvenanceProvenance
Where a number came from.
Attached per probability row rather than per model, and it survives into every answer computed from it. Four tags: elicited from a source, asserted by a named expert, fitted from counts, or mixed. Per-model provenance could not support this, because the average of three kinds of support is the one thing none of them is.
R
- RefutationEstimation
Deliberate attempts to break a result, reported one by one.
The battery runs an unmeasured-confounder stress test, a conditional-association comparison and a fragility bound. Each reports pass, fail or inconclusive, and inconclusive is never counted as a pass — otherwise every unrun check improves the result and the battery always passes. A passing battery means the answer survived these attacks, not that the assumptions hold.
Related: e value, sensitivity envelope
- ReplayRecords
Asking a stored question again and comparing the answer to the one on record.
The comparison is classified as identical, drifted, or reidentified — meaning the question now identifies differently, which is a different failure from a number moving, because the second can be a model update and the first means the question is no longer answerable the same way.
Related: evidence pack, decision record
- Risk differenceEstimation
The absolute change in the probability of an outcome between two treatment levels.
The default effect measure, reported alongside a risk ratio. An absolute change is usually what a decision turns on — a 36 percentage point reduction and a halving are the same ratio and very different decisions — so the difference is the headline and the ratio accompanies it.
Related: estimand, sensitivity envelope
S
- SegmentEstimation
One subgroup's worth of a heterogeneous effect.
Segments are ranked rather than reported in isolation, because the useful question is usually which subgroup to act on first. A continuous segmentation variable has to be binned to be used, and binning is a modeling choice that changes what the answer means.
Related: heterogeneous effect, targeting
- Sensitivity envelopeEstimation
The range an answer moves within when the inputs it rests on are shifted to the ends of their plausible bands.
Every effect carries one. It names its own basis as sensitivity_envelope, ensemble_envelope or combined, and it never means sampling error — an elicited model has no sample to draw error from. A row with no calibrated interval contributes no width rather than an invented one, and where no row is covered there is no band at all.
Related: risk difference, refutation
- StructureStructure
The graph: which variables exist and which ones affect which.
Elicited from content or recovered from conditional independence tests on a table. Each edge carries its source and, where relevant, the contested flag. It is kept separate from the probabilities on purpose, because a graph can be right while every number in it is wrong and the reverse — and merging them makes it impossible to say which is being challenged.
Related: contested, provenance
T
- TargetingEstimation
Choosing which subgroups to act on when the capacity to act is limited.
Ranking and thresholding over per-subgroup effects against a stated capacity. It answers a different and usually better question than propensity modeling: not who is likely to respond, but where an action changes the outcome most per unit of capacity — and why some subgroups should not be targeted at all.
Related: segment, heterogeneous effect
V
- Variable eliminationEstimation
Computing a probability by summing out one variable at a time instead of enumerating every combination.
The inference method. It is exact given the tables, which is why two calls on one revision are byte-identical and there is no seed, no convergence and no Monte Carlo step. Where the widest elimination will not finish, the request is refused with the variable that made it wide rather than allowed to time out.
Related: discrete bayesian network, identification