Unconditional Exchangeability of Treatment Assignment

Why Independence Matters

The difference between the average observed outcomes of treated and untreated people is easy to calculate. Its causal meaning, however, depends on how treatment was assigned. If people who receive treatment would have had different outcomes even without it, the observed difference can mix a treatment effect with pre-existing differences. Unconditional exchangeability is a precise condition under which that problem does not arise for mean outcomes: treatment assignment is independent of the potential outcomes, without conditioning on other variables.

Let $A$ be a binary treatment indicator, with $A=1$ for treatment and $A=0$ for control. For each unit, let $Y(1)$ and $Y(0)$ be the potential outcomes under treatment and control, respectively. The observed outcome is $Y$. We assume the treatment and potential outcomes are well-defined, and that consistency holds: the outcome observed for a unit equals its potential outcome under the treatment actually received.

Definition: Treatment assignment is unconditionally exchangeable with respect to the potential outcomes if $$ A \perp\!\!\!\perp \bigl(Y(0),Y(1)\bigr). $$ That is, the treatment indicator is statistically independent of the joint pair of potential outcomes. Equivalently, the joint distribution of $\bigl(Y(0),Y(1)\bigr)$ is the same among treated and untreated units. This implies that the distribution, and therefore the mean, of each potential outcome is the same in the two treatment groups.

The word “unconditional” matters. The independence statement is made for the population as a whole; it does not condition on covariates. This is a strong assumption in observational studies. In a randomized study, it is supported by the assignment mechanism: treatment is allocated by chance rather than chosen according to units’ potential outcomes. Randomization does not guarantee that a particular sample will have identical characteristics in its two groups, but it makes assignment independent of the potential outcomes under the design.

Identification of the Average Treatment Effect

The average treatment effect (ATE) is the population mean difference between the two potential outcomes:

\mathrm{ATE}=E\bigl[Y(1)-Y(0)\bigr] =E\bigl[Y(1)\bigr]-E\bigl[Y(0)\bigr].

To identify this effect from observed data, we need more than exchangeability. We require consistency, unconditional exchangeability, positive probability of receiving either treatment, and finite means for the outcomes. The positivity condition here is unconditional: $0<P(A=1)<1$. It ensures that both treatment groups exist in the population, so both observed group means are defined.

Identification result: If consistency holds, $A \perp\!\!\!\perp (Y(0),Y(1))$, $0<P(A=1)<1$, and the relevant means are finite, then $$ E[Y\mid A=1]=E[Y(1)], \qquad E[Y\mid A=0]=E[Y(0)]. $$ Consequently, $$ \mathrm{ATE}=E[Y\mid A=1]-E[Y\mid A=0]. $$

The result follows by identifying each potential-outcome mean separately. Exchangeability implies that conditioning on treatment assignment does not change the distribution of either potential outcome. Thus $E[Y(1)\mid A=1]=E[Y(1)]$ and $E[Y(0)\mid A=0]=E[Y(0)]$. Consistency supplies the link to observed outcomes: among treated units, $Y=Y(1)$; among controls, $Y=Y(0)$. Combining these statements gives the result.

\begin{aligned} E[Y\mid A=1] &=E[Y(1)\mid A=1] &&\text{(consistency)}\\ &=E[Y(1)] &&\text{(exchangeability)},\\[4pt] E[Y\mid A=0] &=E[Y(0)\mid A=0] &&\text{(consistency)}\\ &=E[Y(0)] &&\text{(exchangeability)}. \end{aligned}

Subtracting the second equality from the first yields the observed difference-in-means formula for the ATE. In a sample, the corresponding estimate is the sample mean of $Y$ in the treated group minus the sample mean in the control group. The identification result concerns population quantities; sampling variation means that a sample estimate need not equal the population ATE exactly.

Worked Example: A Randomized Treatment and a Mean Difference

In a hypothetical randomized trial, participants are assigned to a program ($A=1$) or usual care ($A=0$). The outcome is a symptom score in points, with higher scores indicating more symptoms. Among 120 participants assigned to the program, the observed mean is 8.6 points. Among 120 assigned to usual care, it is 11.3 points. Suppose the trial uses a well-defined intervention and outcome measurement, and participants’ observed outcomes equal their potential outcomes under the assigned treatment.

Check the identifying conditions. The random assignment mechanism supports independence between assignment and the potential outcomes. Both groups have positive assignment probability, and the specified means are finite. The stated consistency condition links each observed score to the potential outcome for the assigned group. These conditions support identifying the ATE by the difference in population group means; in practice, the reported sample means estimate those population quantities.

Calculate the sample difference.

\bar{Y}_{A=1}-\bar{Y}_{A=0} =8.6-11.3=-2.7\text{ points}.

The estimated average effect is a reduction of 2.7 symptom-score points for the program compared with usual care. This is an average contrast, not a claim that every participant’s score would fall by 2.7 points. The calculation alone does not quantify sampling uncertainty; an interval or test would be needed to assess that uncertainty.

When Unconditional Exchangeability Fails

In an observational study, people may choose treatment, be selected for it, or receive it based on clinical decisions. If those processes are related to potential outcomes, the treated and untreated groups can differ even in the absence of treatment. Then the observed difference in means need not identify the ATE.

Worked Example: Treatment Selection Creates a Misleading Difference

Consider a hypothetical population of 100 people: 50 with high baseline risk and 50 with low baseline risk. The outcome is a risk score in points. Suppose the potential outcomes are as follows: for each high-risk person, $Y(0)=10$ and $Y(1)=8$; for each low-risk person, $Y(0)=4$ and $Y(1)=2$. In each risk group, treatment lowers the score by 2 points. However, 40 of the high-risk people and 10 of the low-risk people receive treatment. Assume the observed outcome equals the potential outcome under the treatment received.

Find the population ATE. Half the population is high risk and half is low risk. The mean effect in each group is $-2$ points, so

\mathrm{ATE} =0.5(8-10)+0.5(2-4) =0.5(-2)+0.5(-2) =-2\text{ points}.

Calculate the observed group means. The treated group contains 40 high-risk people with observed scores of 8 and 10 low-risk people with scores of 2. Its mean is

\bar{Y}_{A=1} =\frac{40(8)+10(2)}{50} =\frac{340}{50} =6.8\text{ points}.

The control group contains 10 high-risk people with scores of 10 and 40 low-risk people with scores of 4. Its mean is

\bar{Y}_{A=0} =\frac{10(10)+40(4)}{50} =\frac{260}{50} =5.2\text{ points}.

The observed difference is $6.8-5.2=1.6$ points, even though the treatment lowers the outcome by 2 points for everyone. The groups have different distributions of baseline risk, so assignment is not independent of the potential outcomes. For example, the treated group has a higher mean $Y(0)$ than the untreated group: $\{40(10)+10(4)\}/50=8.8$ points, compared with $\{10(10)+40(4)\}/50=5.2$ points. Consistency and treatment positivity hold here, but unconditional exchangeability does not. The observed difference therefore does not identify the ATE.

What the Assumption Does and Does Not Say

Unconditional exchangeability is a claim about the assignment process and the potential outcomes, not merely about the observed outcomes. In observed data, each unit reveals only one potential outcome. Consequently, the unobserved potential outcomes cannot be compared directly between the assigned groups to verify independence. Similar observed group averages, or balance on a few recorded characteristics, do not establish independence of the full potential-outcome pair.

The assumption also does not say that treatment has no effect. It says that assignment does not select units according to their potential outcomes. The potential outcomes may differ greatly from one unit to another, and treatment effects may vary; exchangeability concerns whether those potential outcomes are distributed the same way across assignment groups.

Worked Example: Averages Do Not Reveal Every Individual Effect

Imagine a hypothetical population with four equally common potential-outcome types, measured in outcome units: $(Y(0),Y(1))=(0,2),(2,4),(4,3),(6,5)$. Suppose treatment is assigned independently of type, with $P(A=1)=0.5$, and consistency holds.

The mean under control and the mean under treatment are

E[Y(0)]=\frac{0+2+4+6}{4}=3, \qquad E[Y(1)]=\frac{2+4+3+5}{4}=3.5.

Therefore, the ATE is

\mathrm{ATE}=3.5-3=0.5\text{ outcome units}.

The individual effects in the four types are $2,2,-1,-1$ units. Their average is $\{2+2-1-1\}/4=0.5$ units, but the average alone does not disclose those individual effects. With independent assignment, observed treated and control outcomes identify the two marginal means in expectation; they do not reveal both potential outcomes for the same person. Thus identifying an average effect is not the same as identifying each person’s effect or the joint distribution of the two potential outcomes.

Diagnostics, Alternatives, and Limitations

Because one potential outcome is missing for every unit, unconditional exchangeability cannot generally be tested from the observed outcomes. Diagnostics can still identify reasons to doubt the assumption. Investigators should examine how treatment was assigned, whether allocation was genuinely randomized, whether assignment rules used prognostic information, and whether the treatment groups show substantial imbalance in measured pretreatment characteristics. In a randomized study, chance imbalance is possible; in an observational study, imbalance may signal selection. Neither balance nor lack of balance on measured variables proves or disproves independence of the full potential-outcome pair.

If assignment is randomized, the design itself provides the central justification for exchangeability. If treatment is observational and assignment depends on prognosis, an unadjusted difference in means is generally only a descriptive contrast. An analysis that accounts for covariates may be appropriate under a different, conditional exchangeability assumption; that is not the unconditional assumption defined here. Other study designs may also target causal effects, but each requires its own identifying assumptions.

Unconditional exchangeability has important limits. It can be implausible when treatment is selected based on health status, preferences, access, or other prognostic factors. It is also an assumption about the assignment mechanism, not a feature that follows from a large sample. A larger sample can make a mean estimate more precise while leaving confounding intact. Finally, even when the assumption is credible, identification of a population mean does not by itself establish that the sample estimate is precise or practically important.

Common mistake: “The treated and control groups have similar observed outcome averages, so treatment assignment must be exchangeable.” This is not valid: the observed averages are outcomes under different treatments, and the missing potential outcomes cannot be checked from them. Exchangeability concerns the joint potential outcomes and needs support from the assignment design or a defensible substantive argument. Balance on observed characteristics is useful diagnostic information, not proof.

A Practical Identification Check

Define the causal contrast.
State the treatment states, the outcome, and the population mean contrast $E[Y(1)]-E[Y(0)]$.

Assess exchangeability.
Ask whether assignment is independent of the joint potential-outcome pair, and explain why the design makes that credible or why it may fail.

Check the remaining links.
Confirm consistency and that both treatment groups have positive probability, with finite outcome means.

Use the identified contrast carefully.
When the conditions hold, compare the population observed means; describe a sample difference as an estimate and report its uncertainty separately.

Check Your Understanding

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

Write unconditional exchangeability for a binary treatment and its two potential outcomes.
A study has consistency and unconditional exchangeability, but nobody in the population receives control. Can the two-group difference identify the ATE? Explain.
In a hypothetical randomized study, the treated mean is 14 units and the control mean is 17 units. Under the identification conditions, what is the estimated direction and size of the average effect?
Why does balance on measured baseline variables not prove unconditional exchangeability?

Solutions. (1) $A\perp\!\!\!\perp (Y(0),Y(1))$. (2) No. If nobody receives control, $P(A=0)=0$, so the control-group mean is unavailable and the positivity condition fails. (3) The difference is $14-17=-3$ units, an estimated average reduction of 3 units under treatment. (4) The assumption concerns the joint distribution of both potential outcomes, including unobserved outcomes; observed balance only describes measured variables and cannot establish that independence.