Identification
Before the number, check the question.
An effect is identified when the model's structure supports computing it without unmeasured confounding. Many questions are not. This system reports the verdict, the reasoning and the adjustment set — and returns the fact rather than a number when the structure does not support the claim.
The number is the easy part. Whether the question can be answered at all is the part that decides whether the number means anything.
Two different questions
Both get asked, and they have different answers. Treating them as one is the most common way a causal analysis goes wrong.
Is it answerable?
A question about structure. Does the model contain an open path between treatment and outcome that nothing you can measure blocks? If it does, no amount of data helps — a larger sample estimates the wrong quantity more precisely.
Do you have support for it?
A question about data. Even when the structure is fine, if some combination of adjustment variables has no observations, the number rests on extrapolation. That is the positivity check, and it is reported separately.
Did the answer survive being attacked?
A question about robustness. Three refutation checks that can fail, plus a fragility measure. A check that cannot run returns inconclusive and is never counted as a pass.
What a refusal actually looks like
This is unedited output from a stored answer, not an illustration. The question is the textbook reverse-causality case.
- Treatment
- Lung Cancer
- Outcome
- Smoking
an open backdoor path remains: some common cause of the treatment and the outcome is unobserved, so the effect is not identified from observational data without further assumptions
- A number that cannot be defended is worse than no number. No number gets questioned. A plausible number gets signed off, budgeted against, and reported upward — and the failure is not that the analysis was inconclusive, it is that nobody found out until the outcome moved the other way.
- The refusal is specific. It names the open backdoor path rather than reporting insufficient data or a model error, so the next action is visible: the fix is in the model, not in the data.
- It is a result rather than an error. identified: false flows through the same storage, the same provenance and the same record as a positive answer, and it appears in the history beside the answers that worked.
The related checks
Identification is the strongest of these. It is not the only one that can fail.
Is the effect supported by the data?
Identification is a claim about structure; support is a claim about data. The positivity check reports which strata an adjustment set has no observations for, because a dropped stratum and an extrapolated one look identical in a single number.
What this does not establish: A supported stratum is one that has rows. It is not evidence that the model's structure is right.
Is the adjustment set sound?
Adjusting for a variable caused by the treatment moves the estimate away from the truth, and it usually looks like an improvement because the number settles down. The engine reports the warning and names the variable.
What this does not establish: A warning is raised on a criterion, not on a judgment about your domain. A variable it flags may be defensible for reasons the model cannot see.
Did the answer survive being attacked?
Three checks that can fail, plus a fragility measure. A check that cannot fail is decoration, which is why a placebo-treatment check is deliberately not shipped: against a stated model it is a tautology.
What this does not establish: None of this is validation. A battery that passes means the answer survived these attacks, not that the confounding assumption holds.
What this page is not claiming
A page that lists what a system cannot do is more useful than one that lists what it can, because only the first can be checked.
Not an estimate from data
Estimator libraries learn an effect from observations. This is exact arithmetic on a stated model. Exactness is a property of the arithmetic, not evidence that the model is a faithful description of the world.
Not a confidence interval
The band is a sensitivity envelope — how far the number moves as the assumptions move. There is no sample to draw sampling error from, because the model is elicited rather than observed.
Not a proof of non-identifiability
A refusal means this check did not establish the effect under the stated assumptions. It is not a theorem that no method could, and a richer model may identify what this one does not.
Not per-unit effects
Segments are exact arithmetic within a defined subgroup. They are not learned individual treatment effects, and nothing here produces a counterfactual for a specific person.
If you have a decision to make and a number someone will hold you to, the question is not whether we can produce a number. It is whether you can say where it came from.
Book a walkthroughBring a decision. We will walk through the model, its probability sources, the supported analysis and the record.