Why Identification Assumptions Matter
An adjusted regression can describe how outcomes differ between people with different observed exposures. To interpret that difference as a causal effect, however, requires more than fitting a model or adding covariates. The data must stand in for comparisons that would have been made under well-defined alternative exposures. Identification assumptions explain when that substitution is justified.
Let \(A\) be a binary exposure, with \(A=1\) denoting the exposure of interest and \(A=0\) its comparison, and let \(Y\) be an outcome measured after exposure. Let \(Y(a)\) denote the outcome a person would have under exposure level \(a\). For a specified target population, one common causal contrast is the average difference:
Only one of \(Y(1)\) and \(Y(0)\) is observed for each person. In observational data, the people who receive the exposure may also differ from those who do not. Four assumptions—consistency, conditional exchangeability, positivity, and no interference—state what must hold for observed data to identify this contrast. They do not guarantee that a particular regression model estimates it accurately.
Four Assumptions for Identification
- Consistency: a person's observed outcome equals the potential outcome under that person's observed exposure: \(Y=Y(A)\). The exposure must correspond to a sufficiently well-defined intervention; materially different versions should not be treated as one exposure if they can produce different outcomes.
- Conditional exchangeability: for each \(a\in\{0,1\}\), \(Y(a)\perp A\mid L\). Within levels of \(L\), exposure assignment is independent of the potential outcome under \(a\).
- Positivity: for every covariate profile \(l\) occurring in the target population, \(P(A=a\mid L=l)>0\) for each exposure level \(a\) in the contrast.
- No interference: one person's potential outcome does not depend on other people's exposure assignments. In a group of \(n\) people, \(Y_i(a_1,\ldots,a_n)=Y_i(a_i)\).
These assumptions answer different questions. Consistency concerns what the exposure means and how observed outcomes relate to potential outcomes. Exchangeability concerns comparability of exposed and unexposed people after conditioning on \(L\). Positivity concerns whether both exposure levels occur for the covariate profiles being compared. No interference concerns whether one person's exposure can affect another person's outcome.
The covariates in \(L\) should be appropriate for the causal question and measured before exposure. Conditional exchangeability is a statement about potential outcomes, not merely a claim that the observed covariate distributions look similar. In an observational study, the potential outcomes are not jointly observed, so exchangeability cannot generally be verified from the observed data alone. Subject-matter knowledge and a defensible account of exposure assignment are essential.
Consistency: A Clear Exposure and Outcome
Consistency is often written \(Y=Y(A)\): if a person actually receives exposure \(a\), their observed outcome is the outcome they would have had under \(a\). This link is meaningful only if “receives exposure \(a\)” has a sufficiently clear interpretation. If the same exposure label covers versions with importantly different effects, a single potential outcome \(Y(a)\) may not represent a coherent intervention.
Worked Example: Defining a Medication Exposure
Suppose a hypothetical clinic dataset codes \(A=1\) if a patient received “pain medication” before discharge and records whether the patient returned to the emergency department within 7 days. In 40 exposed patients, 6 returned, so the observed return proportion is \(6/40=0.15\), or 15%.
The arithmetic describes the observed group. To use \(Y(1)\) in a causal contrast, the exposure level must be interpretable. If \(A=1\) combines several drugs, doses, and timing schedules that would lead to meaningfully different potential outcomes, “pain medication” may be too vague to define one intervention. A more specific contrast, such as a named regimen versus no medication under a stated protocol, may make consistency more plausible.
If the exposure is sufficiently well-defined, consistency says that each exposed patient's observed return outcome is their \(Y(1)\), and each unexposed patient's observed outcome is their \(Y(0)\). It does not say that the 15% observed return proportion is \(E[Y(1)]\) for the target population: the exposed patients may differ systematically from that population. That further inference requires the other assumptions as well.
Consistency is not a claim that every person responds in the same way. It connects each person's observed outcome with that person's potential outcome under the exposure actually received. Nor does it eliminate the fundamental fact that the alternative potential outcome for that person remains unobserved.
Conditional Exchangeability: Comparable Potential Outcomes
Conditional exchangeability says that, after accounting for \(L\), exposure groups are comparable with respect to each potential outcome. It is a formal expression of the requirement that the measured covariates suffice to address confounding for the chosen contrast. If an unmeasured pre-exposure factor influences both exposure and outcome, the condition may fail even when the fitted model includes many measured variables.
Worked Example: Conditioning on Baseline Severity
Imagine a hypothetical observational comparison of a new treatment (\(A=1\)) with usual care (\(A=0\)) for recovery within 30 days (\(Y=1\) for recovery). Clinicians tend to give the new treatment to patients with severe baseline illness. Baseline severity \(L\) is classified as mild or severe.
Suppose the proportion treated is 70% among patients with severe illness and 30% among those with mild illness. This shows that exposure assignment depends on severity, so an unadjusted comparison mixes treatment differences with severity differences. It does not by itself establish that conditional exchangeability holds or fails.
For the intended adjustment to support identification, the substantive claim would be that, within each severity category, treatment assignment is independent of the potential recovery outcomes: \(Y(a)\perp A\mid L\), for both \(a=0\) and \(a=1\). If other factors—such as a clinician's unrecorded assessment of prognosis—affect both treatment choice and recovery, severity alone may not be sufficient. The conditional exchangeability claim would then lack support, even if severity is included in a regression.
The example separates two ideas: an observed treatment imbalance can motivate adjustment, but the assumption is about potential outcomes and cannot be confirmed just by inspecting that imbalance. The justification depends on whether the measured pre-exposure covariates adequately capture common causes of exposure and outcome.
Conditioning on variables is not automatically beneficial. Covariates should be chosen to make a credible exchangeability argument, not simply because they predict the outcome. The earlier material on confounders, mediators, and colliders helps distinguish relevant pre-exposure adjustment variables from variables that should not be conditioned on for the intended causal question.
Positivity and No Interference
Positivity is specific to the target population and the adjustment covariates. If a covariate profile occurs in the target population but everyone with that profile receives the same exposure, the data provide no observed comparison for the other exposure level at that profile. The relevant condition is positive probability for each exposure level in the contrast at every target-population profile. This tutorial states that requirement; a detailed assessment of covariate overlap is a separate task.
No interference rules out one person's exposure changing another person's outcome. For example, if one patient's treatment alters a shared ward environment in a way that affects other patients' recovery, the individual potential outcome may depend on the treatment assignments of other patients. In that setting, the simple notation \(Y_i(a_i)\) omits part of the intervention. The study may need a group-level or otherwise explicitly defined exposure strategy.
Consistency and no interference are related but distinct. Consistency connects a person's outcome to their own observed exposure under a well-defined exposure description. No interference ensures that other people's exposure assignments do not also determine that outcome. A study can have a clearly specified treatment but still face interference, as may occur when people interact or share resources.
How the Assumptions Identify a Causal Contrast
For a discrete covariate \(L\), under consistency, conditional exchangeability, positivity, and no interference, the mean potential outcome under exposure \(a\) can be written as a standardization formula. The argument uses the law of total expectation and then replaces the unobserved potential-outcome mean with an observed conditional outcome mean.
\(E[Y(a)]=\sum_l E[Y(a)\mid L=l]P(L=l)\). This is the law of total expectation.
Because \(Y(a)\perp A\mid L\), \(E[Y(a)\mid L=l]=E[Y(a)\mid A=a,L=l]\), wherever the conditional exposure group exists.
Among people with \(A=a\), consistency gives \(Y=Y(a)\). Therefore \(E[Y(a)\mid A=a,L=l]=E[Y\mid A=a,L=l]\).
Positivity ensures the conditional observed mean in step 3 is defined for each target profile. No interference supports representing the relevant potential outcome as depending on the person's own exposure rather than the whole exposure pattern.
Combining the steps gives the identifying result:
Thus the population-average contrast is identified by subtracting the standardized mean for \(a=0\) from the standardized mean for \(a=1\). For continuous covariates, the sum is replaced by an integral over the target population's covariate distribution. The assumptions identify the contrast in terms of observed-data quantities; they do not dictate one particular estimation method or guarantee that a chosen statistical model is correct.
Worked Example: Standardizing Recovery Outcomes
Consider hypothetical data on 100 patients. Baseline severity is mild for 60 patients and severe for 40. Within each severity group, both treatment choices are observed. The outcome is recovery within 30 days.
| Baseline severity | New treatment: recovered / total | Usual care: recovered / total |
|---|---|---|
| Mild | 8 / 40 = 0.20 | 2 / 20 = 0.10 |
| Severe | 12 / 20 = 0.60 | 12 / 20 = 0.60 |
Suppose the target population has the observed severity distribution: 60% mild and 40% severe. If all four identification assumptions hold, standardization estimates the mean recovery outcome under each exposure by applying each within-severity outcome rate to that target distribution.
For the new treatment:
For usual care:
The standardized estimated risk difference is \(0.36-0.30=0.06\), or 6 percentage points. Under the assumptions, the estimate represents an average increase of 6 percentage points in 30-day recovery in this target population under the new treatment rather than usual care. It is not a claim that every patient benefits by 6 points, nor does the calculation prove the assumptions. For example, unmeasured prognosis could still make conditional exchangeability implausible.
The example also illustrates positivity: each severity group contains patients under both exposure levels, so all four conditional proportions can be calculated. The calculation alone does not establish that other relevant covariate profiles have adequate representation, or that the exposure has no versions or spillover effects.
What These Assumptions Do—and Do Not—Establish
The assumptions provide a bridge from an observational comparison to a causal estimand, but they are not all equally assessable from the data. Researchers can examine whether the recorded exposure definition is clear, whether both groups occur across observed profiles, and whether interference is plausible given the setting. They must also use subject-matter knowledge to defend the covariate set and the exchangeability claim. No regression output can demonstrate that all unmeasured common causes have been controlled.
Keep identification separate from estimation. Identification asks whether the target causal quantity can, in principle, be expressed using the observed-data distribution under stated assumptions. Estimation asks how to calculate an estimate from a finite dataset, often using a regression model or another method. A well-fitting model cannot repair a failed identification assumption, while a credible identification argument does not guarantee precise estimates.
Common Mistakes
- “We adjusted for several variables, so the effect is causal.” Adjustment supports a causal interpretation only if the covariates are appropriate and conditional exchangeability is credible; the number of variables is not a guarantee.
- “The treatment groups look similar, so exchangeability is proven.” Observed balance cannot establish independence of treatment and unobserved potential outcomes.
- “Both exposure groups exist in the whole sample, so positivity holds.” Positivity must hold within every relevant covariate profile in the target population, not merely overall.
- “Consistency means every patient gets the same outcome.” It means an individual's observed outcome matches their potential outcome under their received exposure, not that outcomes are identical across people.
- “No interference is the same as consistency.” Consistency links an individual's outcome to their own exposure; no interference rules out effects of other people's exposures on that outcome.
- “The standardization calculation verifies the assumptions.” It uses the assumptions to interpret observed conditional means; arithmetic cannot prove the assumptions true.
Check Your Understanding
Answer each question by naming the assumption involved and explaining its role.
- A study combines several substantially different treatment regimens under one exposure label. Which assumption is threatened, and why?
- State conditional exchangeability for a binary exposure \(A\), potential outcome \(Y(a)\), and covariates \(L\).
- Why does observing both treatment groups in the overall dataset fail to establish positivity for every covariate profile?
- In a shared ward, one patient's treatment changes other patients' recovery outcomes. Which assumption is questionable, and what part of the usual individual potential-outcome notation may be inadequate?
- Under what assumptions does standardization identify \(E[Y(a)]\) using \(E[Y\mid A=a,L]\) and the target distribution of \(L\)?