Conditional Exchangeability Across Study Participation

Why Study Participation Needs Its Own Assumption

A treatment effect estimated in a study concerns the people represented by that study. A different population may have different covariate distributions, different potential outcomes within covariate strata, or both. Earlier material on effect modification across populations separates these possibilities: different population averages can arise simply because populations contain different proportions of people with different treatment effects. To learn whether study results can speak to a target population, we also need an assumption about how study participation relates to potential outcomes after accounting for measured covariates.

This assumption concerns who enters the study, not how treatment is assigned among those who enter it. Those are separate mechanisms, and each can threaten a different part of a causal argument.

Definition: Let $S=1$ indicate participation in the study and $S=0$ indicate membership in the target population without study participation. Let $A$ denote treatment, $X$ a set of measured covariates defined for both groups, and $Y(a)$ the potential outcome that would be observed under treatment level $a$. Conditional exchangeability across study participation means that, at each covariate value $x$, the potential outcomes are independent of study participation: $$ \{Y(a):a\in\mathcal{A}\}\ \perp\!\!\!\perp\ S\mid X. $$ A commonly used, weaker version requires $Y(a)\perp\!\!\!\perp S\mid X$ separately for each treatment level $a$. The full-vector version implies this treatment-specific version.

In words, among people with the same measured covariates, study participants and nonparticipants have the same distribution of the potential outcomes under each treatment. The assumption does not say that participants and nonparticipants have the same covariate distribution. It also does not require participation to be unrelated to potential outcomes before conditioning on $X$. Selection may depend strongly on $X$, provided that conditioning on the chosen covariates removes the remaining association with the potential outcomes.

For many questions about mean causal effects, a weaker assumption may be enough: conditional mean exchangeability, meaning $E[Y(a)\mid X=x,S=1]=E[Y(a)\mid X=x,S=0]$ for each relevant $a$ and target covariate value $x$. This equality of means does not imply equality of the full outcome distributions. The full conditional-independence assumption is a stronger statement, so the required version should be matched to the estimand and argument.

Keep Participation and Treatment Assignment Separate

Treatment exchangeability addresses a different comparison. Within the study, it says that treatment assignment is independent of the potential outcomes conditional on measured covariates. For example, a sufficient form for treatment levels $a$ is $Y(a)\perp\!\!\!\perp A\mid X,S=1$. This assumption concerns whether treated and untreated study participants are comparable for causal inference. Conditional exchangeability across participation instead concerns whether study participants and nonparticipants are comparable at the same $X$.

A randomized trial can make treatment assignment independent of potential outcomes among enrolled participants, while enrollment remains related to potential outcomes. Randomization therefore supports a causal comparison within the trial; by itself, it does not make that trial population representative of people who did not enroll.

Assumption	Comparison being made	What it addresses
Treatment exchangeability	Different treatment assignments among study participants, at fixed covariates	Confounding of the treatment comparison
Participation exchangeability	Study participants and nonparticipants, at fixed covariates	Differences in potential outcomes between study and target populations

The covariates needed for these two assumptions need not be identical in every application, although a single set $X$ is used here for clarity. In either case, a covariate set must be defined and measured appropriately for the comparison it is meant to support. A variable measured only among participants cannot, without additional information, directly condition a participant–nonparticipant comparison.

Worked Example: Conditional Mean Exchangeability Despite Different Population Averages

Consider a hypothetical study and target population with a binary covariate $X$. The table gives the mean potential outcomes in each participation group. Assume for this example that the displayed means are the true values in the hypothetical populations.

Covariate stratum	Participation group	Proportion in group	$E[Y(0)]$	$E[Y(1)]$
$X=0$	Study, $S=1$	0.60	2	5
$X=0$	Target, $S=0$	0.40	2	5
$X=1$	Study, $S=1$	0.40	8	10
$X=1$	Target, $S=0$	0.60	8	10

Within $X=0$, the mean under control is 2 in both groups and the mean under treatment is 5 in both groups. Within $X=1$, the corresponding means are 8 and 10 in both groups. Thus the table supports conditional mean exchangeability: the treatment-specific potential outcome means agree across participation groups within each stratum. Means alone would not establish equality of the entire potential outcome distributions.

The conditional treatment effects are $5-2=3$ outcome units for $X=0$, and $10-8=2$ outcome units for $X=1$. The study average treatment effect, using its covariate proportions, is $$ 0.60(3)+0.40(2)=1.80+0.80=2.60\ \text{outcome units}. $$ The target average is $$ 0.40(3)+0.60(2)=1.20+1.20=2.40\ \text{outcome units}. $$ The averages differ because the groups have different proportions in the two strata and the effect differs by $X$, even though the conditional potential outcome means agree. Conditional exchangeability does not require the marginal study and target effects to be equal.

What the Assumption Gives—and What It Does Not

Conditional exchangeability across participation has a direct implication for potential outcome means. Fix a treatment level $a$ and a covariate value $x$ represented in both groups. Under $Y(a)\perp\!\!\!\perp S\mid X$, the conditional distribution of $Y(a)$ is the same for $S=1$ and $S=0$. Equal distributions have equal expectations when those expectations exist. Therefore,

E[Y(a)\mid X=x,S=1] = E[Y(a)\mid X=x,S=0].

This is a statement about potential outcomes, including outcomes that are not observed for a person under the treatment they did not receive. It is not verified merely by comparing observed outcomes between study participants and nonparticipants: those observed outcomes may correspond to different treatments, and potential outcomes are not all observed for any one person.

Nor does participation exchangeability alone identify the study’s potential outcome means from observed study data. To connect an observed outcome to a potential outcome, one also needs consistency—if a person receives $A=a$, their observed outcome equals $Y(a)$—and a suitable treatment exchangeability assumption within the study. Treatment positivity is also needed so that treatment level $a$ occurs among participants at the covariate values of interest. For comparisons with a target population, there must also be study representation at those covariate values; otherwise a participant group cannot directly inform that part of the target.

Important distinction: Conditional exchangeability across participation is an assumption, not a consequence of randomization within a study. It is also distinct from overlap. Exchangeability concerns whether potential outcomes differ by participation after conditioning on $X$; overlap concerns whether the study includes participants at the target covariate values needed for a comparison.

Worked Example: A Randomized Study Can Still Have Selection Differences

Suppose a hypothetical trial randomizes treatment among its 100 participants: 50 receive treatment and 50 receive control. Imagine that treatment assignment is independent of the potential outcomes within the study, and that the mean observed outcome is 6 units in the treated group and 4 units in the control group. The randomized contrast is $6-4=2$ outcome units for the study participants.

Now suppose a measured covariate $X$ is the same for everyone in this simplified example, but the potential outcome means differ between participants and nonparticipants:

Group	$E[Y(0)]$	$E[Y(1)]$	Mean treatment effect
Study participants, $S=1$	4	6	$6-4=2$
Nonparticipants, $S=0$	8	10	$10-8=2$

Randomization supports treatment exchangeability in the study, and the observed contrast of 2 units matches the study’s mean potential outcome contrast. But participation exchangeability fails: $E[Y(0)\mid S=1]=4\ne8=E[Y(0)\mid S=0]$, and $E[Y(1)\mid S=1]=6\ne10=E[Y(1)\mid S=0]$. Thus randomization has not made the two groups’ potential outcome levels alike.

In this constructed example, the conditional mean treatment effects happen to agree at 2 units in both groups. That equality is weaker than participation exchangeability for the potential outcome means: both potential outcome means shift by 4 units, even though their difference does not. A question about transporting a mean effect may require an assumption about conditional effects rather than the stronger equality of each treatment-specific potential outcome mean. The assumptions must be stated in terms of the actual target quantity; equality of effects should not be mislabeled as equality of potential outcomes.

Measured Covariates Must Capture the Relevant Selection Differences

The phrase “conditional on measured covariates” is substantive. It is not enough to include convenient predictors of participation. The covariates must suffice to remove differences in the relevant potential outcomes between participants and nonparticipants, at least to the degree required by the estimand. If an unmeasured prognostic factor affects both study participation and potential outcomes within levels of $X$, the conditional independence assumption can fail.

A related possibility is that an omitted covariate modifies the treatment effect. Then participants and nonparticipants can have different treatment effects within the recorded $X$ strata. This is why effect modifiers can matter for generalizing a causal effect: conditioning only on variables that predict participation may not be enough if relevant effect variation remains within those strata.

Worked Example: An Unmeasured Prognostic Factor Violates the Assumption

Suppose everyone has the same recorded value of $X$, but an unmeasured binary factor $U$ is associated with participation. In a hypothetical population, $U=1$ occurs among 80% of study participants and 20% of nonparticipants. Under control, let $Y(0)=10$ when $U=1$ and $Y(0)=2$ when $U=0$. Under treatment, let $Y(1)=7$ when $U=1$ and $Y(1)=1$ when $U=0$.

Among participants, the mean potential outcome under control is $$ E[Y(0)\mid S=1] =0.80(10)+0.20(2) =8.00+0.40 =8.40\ \text{outcome units}. $$ Among nonparticipants, it is $$ E[Y(0)\mid S=0] =0.20(10)+0.80(2) =2.00+1.60 =3.60\ \text{outcome units}. $$ The conditional means differ even though everyone has the same measured $X$, so conditioning on $X$ has not made participation exchangeable for $Y(0)$.

The same calculation under treatment gives $0.80(7)+0.20(1)=5.80$ outcome units among participants and $0.20(7)+0.80(1)=2.20$ among nonparticipants. The conditional mean treatment effects are $5.80-8.40=-2.60$ units for participants and $2.20-3.60=-1.40$ units for nonparticipants. In this example, selection on $U$ produces both different potential outcome levels and different mean effects. Measuring and appropriately conditioning on $U$, if feasible and valid, could address this particular source of difference; merely observing that the recorded $X$ is balanced cannot.

Scope, Overlap, and Assessment

Participation exchangeability is not generally testable from observed data alone. For each person, only the outcome under the received treatment is observed, and nonparticipants may have no study outcomes at all. Researchers can examine measured covariate distributions, study recruitment, and subject-matter knowledge to assess whether the assumption is plausible, but similar observed covariate distributions do not prove equality of unobserved potential outcomes.

The assumption is meaningful only for covariate values where both groups can be considered. If a target stratum contains no study participants, there is no direct within-stratum study information with which to compare potential outcomes. This is a failure of overlap, not evidence for or against exchangeability. Even where overlap holds, unmeasured selection factors can still make exchangeability implausible.

A careful analysis therefore names the participation indicator, target group, potential outcomes, and covariates in the assumption. It also states whether the claim is full conditional independence, treatment-specific independence, equality of conditional means, or merely equality of conditional treatment effects. These claims are not interchangeable.

Common Mistakes

Using randomization as a transportability argument. Randomization addresses treatment assignment among enrolled participants; it does not establish that participants and nonparticipants have comparable potential outcomes.
Calling covariate balance proof of exchangeability. Balance concerns measured covariates. Potential outcome independence is not directly observed and can fail because of unmeasured factors.
Confusing equal effects with equal potential outcomes. Two groups can have the same treatment effect while both potential outcome means differ, as in the randomized-study example.
Ignoring overlap. Exchangeability does not create study information for target covariate strata with no participants.
Leaving the conditioning set vague. State which covariates are conditioned on and why they are relevant to potential outcomes and participation.

Key takeaway: Conditional exchangeability across study participation says that, within levels of measured covariates, potential outcomes do not depend on study participation. It is separate from treatment exchangeability, and neither assumption should be substituted for the other.

Check Your Understanding

Use the definitions and distinctions above to answer these questions.

In one sentence, state what $Y(a)\perp\!\!\!\perp S\mid X$ means in a study and target-population comparison.
Why does random treatment assignment among study participants fail, by itself, to establish exchangeability across participation?
Can conditional mean treatment effects agree across participants and nonparticipants even when the means of $Y(0)$ differ? Explain.
What is the difference between a failure of participation exchangeability and a lack of study overlap at a target covariate value?
Why can observed balance in measured covariates not prove conditional independence of potential outcomes and study participation?

Covariate stratum	Participation group	Proportion in group	\(E[Y(0)]\)	\(E[Y(1)]\)
\(X=0\)	Study, \(S=1\)	0.60	2	5
\(X=0\)	Target, \(S=0\)	0.40	2	5
\(X=1\)	Study, \(S=1\)	0.40	8	10
\(X=1\)	Target, \(S=0\)	0.60	8	10

Group	\(E[Y(0)]\)	\(E[Y(1)]\)	Mean treatment effect
Study participants, \(S=1\)	4	6	\(6-4=2\)
Nonparticipants, \(S=0\)	8	10	\(10-8=2\)