Conditional Exchangeability Given Covariates

Why Compare Treatment Groups Within Covariate Strata?

In an observational study, people who receive treatment may differ systematically from those who do not. A difference in their observed outcomes can therefore reflect both treatment and pre-existing differences. Pretreatment covariates offer a way to define groups of people who share measured characteristics and compare treated and untreated people within those covariate strata. The key question is whether those measured covariates suffice to make treatment assignment unrelated to the potential outcomes within each stratum.

The earlier material on unconditional exchangeability considered the stronger condition that treatment assignment is unrelated to potential outcomes overall. Conditional exchangeability relaxes that requirement: treatment groups need not be comparable in the full population, provided they are comparable after conditioning on the relevant pretreatment covariates. This is an assumption about how treatment was assigned, not a property established merely by calculating adjusted outcome differences.

Definition: Let $A\in\{0,1\}$ denote treatment, $Y(1)$ and $Y(0)$ the potential outcomes under treatment and control, and $X$ a set of pretreatment covariates. Conditional exchangeability given $X$ means $$ (Y(0),Y(1))\perp A\mid X. $$ That is, among units with the same covariate value $X=x$, treatment assignment carries no further information about their potential outcomes. For mean-effect arguments, the corresponding condition on conditional means is sufficient: $E[Y(a)\mid A=a,X=x]=E[Y(a)\mid X=x]$ for each treatment state $a$ where the treated-state conditional mean is defined.

Here $X$ must be measured before treatment. It may contain one or several characteristics, such as baseline disease severity, age, or prior outcome history. Conditioning on $X$ means comparing units at the same value—or, for continuous covariates, the same covariate profile in the relevant conditional-distribution sense. It does not mean that every unit has an identical counterpart, nor does it guarantee that all important causes of both treatment and outcome have been measured.

What the Assumption Lets Us Say

Consistency links the observed outcome to the potential outcome corresponding to the treatment actually received: when $A=a$, $Y=Y(a)$. Conditional exchangeability then links the potential outcomes of those receiving state $a$ to the potential outcomes of all units at the same $X=x$. That link is available only where state $a$ occurs with positive probability. For a comparison of both states at $x$, both treatment probabilities must be positive.

\begin{aligned} E[Y\mid A=a,X=x] &=E[Y(a)\mid A=a,X=x] \\ &=E[Y(a)\mid X=x], \qquad \text{when }P(A=a\mid X=x)>0. \end{aligned}

The first equality follows from consistency for units with $A=a$. The second follows from conditional exchangeability, together with the condition that the treated-state conditional mean exists at that covariate value. If both treatment states are represented at $x$, the reasoning applies separately to $a=0$ and $a=1$. Subtracting the two conditional means then gives the mean causal effect among units with $X=x$:

E[Y(1)-Y(0)\mid X=x] = E[Y\mid A=1,X=x]-E[Y\mid A=0,X=x], \qquad 0<P(A=1\mid X=x)<1.

This is a within-stratum identification statement: under the stated assumptions and support condition, the observed treatment-group contrast at $x$ equals the causal contrast at $x$. It does not say that the overall, unadjusted difference in observed means is causal. Nor does it specify how to combine stratum-specific effects into a population-level effect; that requires defining the target population and an appropriate averaging step.

Worked Example: Within-Stratum Comparisons Can Differ From the Crude Difference

Suppose a hypothetical study compares a treatment for reducing a symptom score, where lower scores are better. Baseline symptom severity $X$ has two categories, low and high. The observed group sizes and mean follow-up scores are:

Baseline severity	Treated: number	Treated: mean score	Control: number	Control: mean score
Low	40	8	60	10
High	60	18	40	20

There are treated and control participants in each severity stratum, so both observed conditional means can be calculated. The within-stratum differences are:

\begin{aligned} \text{Low severity: }&8-10=-2\text{ score points},\\ \text{High severity: }&18-20=-2\text{ score points}. \end{aligned}

If consistency and conditional exchangeability given baseline severity hold, these differences identify the mean causal effect within their respective strata. In context, the treatment reduces the mean score by 2 points within each severity group. The assumption is substantive; the table alone does not demonstrate that it holds.

The crude difference is different. The treated mean is $(40\times8+60\times18)/100=14$ points, while the control mean is $(60\times10+40\times20)/100=14$ points. Thus the crude difference is $14-14=0$ points, even though each within-stratum contrast is $-2$. Treated participants were more often in the high-severity group, whose scores are higher in both treatment groups. The example shows why conditional comparability can matter; it does not establish that the assumption is true in any actual study.

How Conditional Exchangeability Can Fail

The assumption says that after conditioning on $X$, the potential outcomes are independent of treatment assignment. It can fail if a prognostic factor that also affects treatment remains unmeasured, is measured poorly, or is not adequately represented by the chosen covariates. In that case, treated and untreated units can still have different potential outcomes within a stratum. A conditional comparison may then mix treatment effects with residual differences in prognosis.

Worked Example: An Unmeasured Prognostic Factor Within a Stratum

Consider a hypothetical study in which all participants have the same recorded baseline covariate $X$, but an unrecorded factor $U$ indicates high or low prognosis. Suppose the potential symptom scores are:

Prognosis $U$	$Y(0)$	$Y(1)$	Treatment assignment in this example
High risk	12	10	Treated
Low risk	8	6	Control

For either prognosis group, the treatment effect is $Y(1)-Y(0)=-2$ score points. But, by the stated assignment pattern, the treated group consists of high-risk participants and the control group of low-risk participants. The observed treated mean is therefore $10$, and the observed control mean is $8$. The observed conditional difference given the recorded $X$ is $10-8=2$ points, suggesting a higher score under treatment, although the treatment reduces the score by 2 points for each type of participant.

The recorded covariate $X$ does not make assignment exchangeable because $U$ still predicts both treatment assignment and potential outcomes within the $X$ stratum. Conditioning on $X$ alone does not remove that imbalance. If $U$ were measured and included in the conditioning set, exchangeability might become more plausible, but whether it does is a substantive question rather than a numerical consequence of adding a variable.

Support, Assessment, and Limits

The comparison in the definition requires data from the treatment state being compared at the covariate value in question. For both potential-outcome means to be connected to observed conditional means, the treatment probability must satisfy $0<P(A=1\mid X=x)<1$. If a state never occurs at $x$, its conditional observed outcome mean is not available there. Conditional exchangeability alone does not supply that missing mean.

Worked Example: A Stratum With No Observed Controls

In a hypothetical study, all 30 participants with low baseline severity receive treatment. Their mean follow-up score is 6 points. Among participants with high baseline severity, 20 receive treatment with a mean score of 12 points, and 20 receive control with a mean score of 15 points.

For high severity, both states are represented, and the observed conditional difference is $12-15=-3$ score points. Under consistency and conditional exchangeability at that covariate value, this identifies the mean effect for the high-severity stratum.

For low severity, the treated mean is observed, but there are no controls. Thus $P(A=0\mid X=\text{low})=0$, and $E[Y\mid A=0,X=\text{low}]$ is undefined from these data. It would be incorrect to write that this observed mean equals $E[Y(0)\mid X=\text{low}]$, or to calculate a within-stratum treatment contrast. Estimating the missing quantity would require additional information or modeling assumptions; it is not justified by conditional exchangeability alone.

Conditional exchangeability is not directly testable from observed outcomes: for each person, only the potential outcome matching the received treatment is observed. Researchers can assess whether relevant pretreatment covariates were measured and whether treatment assignment plausibly depends on unmeasured prognostic factors, using study design knowledge and subject-matter reasoning. They can also inspect whether both treatment states occur across the covariate profiles being compared. Such checks can reveal problems, but a good-looking observed balance or overlap pattern cannot prove exchangeability.

Compared with relying on unconditional exchangeability, conditioning on pretreatment covariates permits treatment assignment to depend on those covariates, as long as potential outcomes are independent of assignment within covariate strata. Compared with an unadjusted outcome comparison, it targets like-with-like comparisons rather than allowing the groups' covariate composition to drive the contrast. These are differences in the assumptions and comparisons involved, not guarantees that one approach is correct in a particular application.

The assumption has important limits. It may be implausible when treatment decisions use clinical judgment or other factors absent from the data; it may also be undermined by measurement error in recorded covariates. Including a covariate merely because it is available does not ensure that the important sources of confounding have been controlled. The covariates must be pretreatment characteristics relevant to the assignment-outcome relationship, and the causal interpretation remains conditional on the assumptions.

Common mistake: “After adjusting for the observed covariates, treatment and control are exchangeable.” Adjustment is a method of making comparisons, not proof of the assumption. The correction is to state conditional exchangeability as a substantive condition, explain why the chosen pretreatment covariates plausibly suffice, and check that both treatment states are represented in the strata being compared.

A Practical Reasoning Sequence

When using conditional exchangeability to interpret an observed comparison, keep the logic explicit:

Name the covariates and timing.
Specify $X$ and establish that its components are measured before treatment.

State the exchangeability claim.
Ask whether, within $X$ strata, treatment assignment is independent of both potential outcomes.

Check the treatment state being compared.
Use the observed conditional mean for state $a$ only where $P(A=a\mid X=x)>0$; a two-state contrast requires both probabilities to be positive.

Interpret only the supported conditional contrast.
With consistency and exchangeability, the observed difference within a supported stratum identifies its mean causal contrast; it does not automatically identify every population-level effect.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

In your own words, what does $(Y(0),Y(1))\perp A\mid X$ assert?
At a covariate value $x$, suppose $P(A=1\mid X=x)>0$ but $P(A=0\mid X=x)=0$. Which conditional outcome mean is unavailable, and why can no two-state contrast be calculated there?
A study compares treated and control participants within recorded baseline-risk groups, but a clinician’s unrecorded prognosis judgment affects both treatment and outcome. What assumption is threatened?
At one supported covariate stratum, the treated mean outcome is 7 units and the control mean is 9 units. Under consistency and conditional exchangeability, what is the identified conditional mean effect, and how should it be interpreted?

Solutions. (1) Within each covariate stratum, treatment assignment carries no information about either potential outcome. (2) $E[Y\mid A=0,X=x]$ is unavailable because no controls occur at $x$; without both observed conditional means, a treatment-control contrast cannot be formed. (3) Conditional exchangeability given the recorded covariates is threatened, because the unrecorded judgment can leave treated and control participants with different potential outcomes within strata. (4) The contrast is $7-9=-2$ units. Under the assumptions, it identifies a 2-unit lower mean outcome under treatment than control among units in that stratum; it is not automatically an effect for the entire population.