When a causal effect is identifiable
A causal effect is identifiable when stated assumptions let you express the causal question as a quantity determined by the observed data.
Fit
When it matters, and when it does not
It matters when
- You want to interpret a treatment or exposure contrast as an effect of changing that exposure.
- You need to explain which assumptions a causal estimate relies on before fitting a model.
It matters less when
- You only want to describe how common an outcome is. Use descriptive statistics without a causal claim. Read: Descriptive statistics and Table 1
- You want to predict an outcome, not change its cause. Use prediction-model guidance. Read: Prediction model development
Before you start
What to check in your study
- A well-defined intervention and causal contrasttargetsummaryMeasureSpecify treatment versions, population, comparator, outcome and time horizon.
- The proposed causal structure and time zeroUse subject knowledge to identify common causes, mediators, selection and spillover between people.
- Data that could support the comparisonCheck whether both treatment options occur across the covariate histories needed for the target population.
- Observed outcomes match the potential outcomes for the treatment actually received.How to check: Potential outcomes describe what would happen under each specified treatment. Define relevant treatment versions, timing and adherence.
- Within adjustment groups, treatment carries no further information about potential outcomes.How to check: This is conditional exchangeability. Justify measured common causes using subject knowledge; balance cannot establish it.
- Both compared treatments are possible in every relevant adjustment group.How to check: Look for empty groups and extreme weights. A model cannot create a treatment option that never occurs.
- The target specifies whether one person can affect another.How to check: No interference is an assumption, not a default truth. With spillover, define group treatment patterns and an appropriate design.
- Selection, missing data and censoring have explicit assumptions.How to check: State the assumptions behind adjustment or weighting. Test sensitivity to plausible unmeasured causes of selection and missing outcomes.
Plain language
How it works
Identification connects the causal target to observed data under assumptions. Estimation uses a finite sample to calculate that quantity, with uncertainty. Even perfect estimation cannot rescue a target the data and assumptions do not identify.
Hill proposed strength, consistency across studies, specificity, temporality, biological gradient, plausibility, coherence, experiment and analogy. They guide judgment across evidence, not a proof checklist. A cause must precede its outcome. Weak effects, several causes or no clear gradient do not rule out causation. Consistency across studies differs from consistency of potential outcomes. Hill considerations cannot establish exchangeability or positivity.
Worked example
A worked example
- Setting
- A clinic compares a new drug with usual care to prevent admission.
- Question
- Is a lower admission risk enough to identify a treatment effect?
- Approach
- Imagine baseline severity and all other common causes are measured, both treatments occur within each group, and follow-up is complete.
- Result
- The standardized risks are 10 in 100 on the drug and 15 in 100 on usual care: a difference of -5 per 100. The risk ratio is 0.67.
- What it means
- A causal reading still needs consistency, exchangeability and the stated spillover assumptions; a balanced dataset cannot prove them.
In practice
Recent studies PubMed search · not vetted
Watch out
Common mistakes
The one to remember
Balance, overlap and a convincing story cannot prove absence of unmeasured confounding. Hill considerations guide judgment across evidence; they are not a checklist that proves causation or identifies a causal effect.
- Treating a balanced covariate table as proof of exchangeability.Instead: Balance concerns measured covariates. Explain unmeasured common causes and assess sensitivity.
- Using a large sample to repair absent treatment overlap.Instead: Restrict the target transparently or redesign the study. Extrapolation adds assumptions and may change the question.
- Counting Hill considerations as a pass-fail causal score.Instead: Weigh the evidence and alternatives. Temporality is necessary, while the other considerations are not universal requirements.
Inside Mallard
How Mallard uses this
Plans do not link to this concept yet. Mallard blocks a plan that lists the same variable as both confounder and mediator. That check cannot establish causal identification.
concept.causal_identification, the id a plan cue records.Related Learn pages
Related designs and methods
Evidence
Sources & further reading
- THE ENVIRONMENT AND DISEASE: ASSOCIATION OR CAUSATION? (opens in a new tab) · Proc R Soc Med · doi 10.1177/003591576505800503
- Estimating causal effects from epidemiological data (opens in a new tab) · J Epidemiol Community Health · doi 10.1136/jech.2004.029496
- Toward Causal Inference With Interference (opens in a new tab) · J Am Stat Assoc · doi 10.1198/016214508000000292