Why Adjustment Has a Causal Goal
In observational data, people who receive a treatment may differ before treatment from people who do not. A comparison of their outcomes can therefore reflect both a treatment effect and pre-existing differences. Adjustment aims to compare treated and untreated people who have the same values of measured, pretreatment covariates. The causal justification for that comparison is conditional exchangeability: after conditioning on those covariates, treatment assignment carries no further information about the potential outcomes.
This is a claim about how treatment groups relate to outcomes that could have occurred, not simply a statement that the groups look similar in the data. It is also an assumption about the assignment process; a statistical calculation cannot establish it from observed outcomes alone.
The full independence statement concerns the joint distribution of both potential outcomes. The mean version concerns only their conditional means, and is sufficient for identifying average outcome contrasts. Neither version says that treated and untreated people have equal outcomes: if treatment has an effect, their potential outcomes can differ. The claim is that, at a given \(X\), the people assigned to each treatment group are comparable in the relevant potential outcomes.
Assumptions That Make the Comparison Causal
Conditional exchangeability is the central goal of adjustment, but it is not the only condition needed to interpret an adjusted comparison causally. The potential outcomes must refer to well-defined treatment options, and the observed outcome must agree with the potential outcome corresponding to the treatment actually received. In addition, both treatment options must be possible in the covariate groups to which the conclusion is meant to apply.
Positivity matters because a comparison cannot be made within a covariate group if that group contains only treated people or only untreated people. Conditional exchangeability matters because even when both groups are present, they may still differ in causes of the outcome that are not adequately represented by \(X\). The earlier material on positivity and on consistency, exchangeability, and positivity treats these conditions separately; here the focus is the role of exchangeability in making like-with-like comparisons informative.
When all conditions hold, the observed conditional mean in treatment group \(a\) identifies the conditional mean potential outcome under \(a\). For a covariate value \(x\), the argument is:
The first equality follows from conditional exchangeability, in its mean form. The second follows from consistency: among people with \(A=a\), the observed outcome is \(Y(a)\). Positivity ensures the conditional mean on the right can be estimated from observed people with \(A=a\) at \(X=x\). Thus, for a target population whose covariate distribution is represented by \(X\), the average treatment effect can be written as
This expression makes the goal clear: estimate the treatment contrast among people with the same measured covariates, then average those conditional contrasts over the target covariate distribution. It does not say that any particular adjustment procedure automatically achieves the goal. The causal interpretation depends on whether the assumptions are credible for the data and population under study.
Worked Comparisons Within Covariate Values
Worked Example: A Conditional Contrast at the Same Baseline Value
Suppose a hypothetical study compares a treatment for a chronic condition with no treatment. Let \(X\) be a pretreatment symptom score, and let \(Y\) be symptom improvement after treatment, measured in points. Among people with \(X=6\), the observed mean improvement is 12.4 points for treated people and 9.1 points for untreated people. What does the within-\(X\) contrast say?
The observed contrast is
This is an observed conditional association: in the study, treated people with a baseline score of 6 improved by 3.3 points more on average than untreated people with that same score. If consistency holds, both treatment options occur among people with \(X=6\), and conditional exchangeability holds there, then the same calculation identifies \(E[Y(1)-Y(0)\mid X=6]\). Without those assumptions, it does not establish that treatment caused a 3.3-point improvement.
The numerical calculation is fully determined by the observed means; whether its causal interpretation is justified is a separate question. In particular, the study must support the claim that relevant differences in potential outcomes have been accounted for at \(X=6\), not merely that the baseline score matches.
Worked Example: Why a Measured Covariate Can Matter
Consider hypothetical observational data in which \(X\) is a pretreatment severity measure. For people with \(X=5\), 70% receive treatment, while for people with \(X=9\), 30% receive treatment. Suppose that, at \(X=5\), the observed mean outcome is 11 units among treated people and 8 units among untreated people. At \(X=9\), the corresponding observed means are 16 and 12 units.
At \(X=5\), the within-value contrast is \(11-8=3\) outcome units. At \(X=9\), it is \(16-12=4\) outcome units. Both covariate values have people in each treatment group, so positivity is satisfied for these two values. If consistency and conditional exchangeability also hold, these conditional differences have causal interpretations for people at the corresponding severity levels.
The unequal treatment frequencies show why an unadjusted comparison could mix treatment contrasts with differences in severity composition. Comparing within a fixed \(X\) removes that measured difference from the particular comparison. It does not establish that severity is the only relevant common cause, or that the exchangeability assumption is true. Those are substantive claims about the assignment process and the chosen covariates.
Worked Example: A Within-Covariate Difference That Is Still Biased
Suppose, hypothetically, that within a fixed measured covariate value \(X=x\), the true mean treatment effect is 2 outcome units for everyone. However, treatment selection also depends on an unrecorded factor. Among treated people, the mean outcome they would have had without treatment is 10; among untreated people, the mean outcome they do have without treatment is 7.
By the stated treatment effect, the treated group's observed mean is \(10+2=12\). The untreated group's observed mean is 7. The observed within-\(X\) contrast is therefore
The causal effect in this example is 2 units, not 5. The other 3 units arise because the treated and untreated groups differ in their mean no-treatment potential outcomes: \(10-7=3\). Conditional exchangeability fails because, even at the same measured \(X\), treatment assignment remains associated with a potential outcome. This example shows why matching measured covariates is not itself proof of causal comparability.
What the Assumption Means in Practice
The substantive question is whether \(X\) captures the pretreatment differences that make treatment selection informative about potential outcomes. If it does, then after fixing \(X\), knowing who received treatment gives no further information about their potential outcomes, as required by exchangeability. If it does not, the conditional comparison may remain confounded.
Exchangeability is not directly testable from observed data: for each person, only one of \(Y(1)\) and \(Y(0)\) is observed. A dataset cannot reveal the missing potential outcome for that person, so it cannot directly confirm that treatment groups had comparable potential outcomes within \(X\). Investigators instead use knowledge of the treatment process, measurement quality, and the study design to judge whether the assumption is credible. A good-looking adjusted comparison is not a substitute for that argument.
Diagnostics can still reveal problems with the proposed comparison. Check whether both treatment groups are represented across the covariate values being compared; an absent group signals a positivity problem. Examine whether the measured pretreatment covariates are comparable after the planned adjustment, and whether important covariate values are missing or recorded unreliably. These checks can expose a failure or weakness in the design, but passing them does not prove exchangeability: observed balance concerns measured covariates, whereas exchangeability concerns potential outcomes.
The main alternatives clarify what adjustment contributes. In a randomized trial, treatment assignment is determined by a random mechanism, so it is independent of potential outcomes by design, possibly subject to chance imbalance in a particular sample. In an observational study, the unadjusted mean difference compares groups as selected; it has a causal interpretation only under stronger conditions, such as unconditional exchangeability. Adjustment instead seeks conditional exchangeability after accounting for \(X\). It can make the comparison more credible when measured covariates explain treatment selection, but it does not turn observational assignment into randomization.
The method also has limits. The covariates used for the comparison must be measured before treatment and must represent the relevant sources of treatment-outcome differences; no amount of calculation can compensate for a covariate set that fails to support exchangeability. Even when the assumption is plausible, limited overlap can restrict which covariate groups support comparisons. Finally, conditional effects may differ across values of \(X\), so a single within-value contrast should not automatically be presented as the effect for every person or for the whole population.
Common Mistake
The practical objective is therefore not to declare that adjustment has removed all confounding. It is to make an explicit, defensible argument that, within the measured pretreatment covariates, treatment groups are exchangeable for the potential outcomes relevant to the causal question. The earlier material on observational identification provides the broader assumption-to-contrast framework; conditional exchangeability is the key assignment assumption within that framework.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- State conditional exchangeability using potential outcomes, treatment assignment, and measured covariates.
- At \(X=x\), the observed treated mean is 14 points and the untreated mean is 10 points. What is the observed conditional contrast, and under what conditions may it be interpreted causally?
- Why does observing both treated and untreated people at each covariate value not, by itself, establish exchangeability?
- Which assumption is directly contradicted if no untreated people are observed at a covariate value included in the target comparison?
- Why can measured-covariate balance after adjustment not prove that the adjusted contrast is causal?
Solutions. (1) Conditional exchangeability is \(\{Y(0),Y(1)\}\perp A\mid X\); for mean effects, the corresponding conditional-mean equality for each treatment level is sufficient. (2) The contrast is \(14-10=4\) points. It identifies a conditional causal effect if consistency and conditional exchangeability hold and both treatment levels are represented at \(X=x\). (3) Presence of both groups supports positivity, but the groups may still differ in potential outcomes within that covariate value. (4) Positivity fails for the no-treatment level at that value, so its conditional mean cannot be estimated there from the observed data. (5) Balance documents similarity in the measured covariates, not independence of treatment assignment and potential outcomes; that independence remains a substantive assumption.