Sequential Exchangeability at Each Decision Time

Why Exchangeability Must Be Sequential

In a longitudinal study, treatment is chosen more than once. A patient’s evolving health may influence the next treatment choice, while also predicting the outcome. The relevant question is therefore not only whether initial treatment is unconfounded, but whether treatment is unconfounded at each decision time, given the information available then. Sequential exchangeability formalizes that requirement.

Let decisions occur at times $t=0,\ldots,K$. Let $L_t$ denote covariates measured before treatment $A_t$, with $L_0$ baseline covariates, and let $Y$ be the final outcome measured after the last decision. Write $\bar A_t=(A_0,\ldots,A_t)$ and $\bar L_t=(L_0,\ldots,L_t)$. The observed history just before decision $t$ is

H_t=(\bar L_t,\bar A_{t-1}),

where $\bar A_{-1}$ is empty, so $H_0=L_0$. This timing convention matters: $L_t$ is available to guide $A_t$, while $L_{t+1}$ occurs after $A_t$. Let $Y^{\bar a_K}$ denote the outcome that would be observed under treatment history $\bar a_K$, and let $L_{t+1}^{\bar a_t}$ denote the next covariate under treatment history $\bar a_t$. Consistency and the meaning of treatment versions are understood as established in the earlier material on those subjects.

Definition: Sequential exchangeability holds if, at every decision time $t$, treatment assignment $A_t$ is conditionally independent of the relevant future potential covariates and outcomes, given the observed history $H_t$. More precisely, for each past treatment history and each continuation of treatment values, the collection of potential covariates after time $t$ and potential outcomes under those continuations is independent of $A_t$, conditional on $H_t$, among histories with that past treatment history. The condition is required only on histories that occur with positive probability.

In compact notation this is often written $A_t\perp\!\!\!\perp\mathcal Z_t\mid H_t$, where $\mathcal Z_t$ collects the relevant counterfactual future quantities. It is a separate condition at every $t$, not a claim that the entire treatment sequence is randomized at baseline. The treatment may depend strongly on $H_t$; exchangeability says that, after conditioning on that history, the assignment contains no additional information about the relevant counterfactual future.

A Local Consequence for Covariate Transitions

A useful way to see the role of the assumption is to examine the covariate measured immediately after a decision. Fix a past history $\bar a_{t-1}$, a current treatment value $a_t$, and a history value $h$ consistent with that past treatment. Under sequential exchangeability, the distribution of $L_{t+1}^{\bar a_t}$ does not differ according to whether the observed treatment at time $t$ was $a_t$. Consistency then connects that counterfactual covariate to the observed next covariate among people who received the specified past and current treatments.

Result: Suppose sequential exchangeability holds for $L_{t+1}^{\bar a_t}$, and consistency holds for the covariate measured after decision $t$. For histories $h$ with positive probability and $P(A_t=a_t\mid H_t=h)>0$, $$ P(L_{t+1}=l\mid H_t=h,A_t=a_t) = P(L_{t+1}^{\bar a_t}=l\mid H_t=h). $$

Proof. By exchangeability, conditional on $H_t=h$, the distribution of $L_{t+1}^{\bar a_t}$ is unchanged by conditioning further on $A_t=a_t$. Thus

P(L_{t+1}^{\bar a_t}=l\mid H_t=h,A_t=a_t) = P(L_{t+1}^{\bar a_t}=l\mid H_t=h).

On the event that the observed past treatments equal $\bar a_{t-1}$ and $A_t=a_t$, consistency gives $L_{t+1}=L_{t+1}^{\bar a_t}$. The history $h$ already includes the specified observed past treatments. Therefore the left side of the preceding equality is $P(L_{t+1}=l\mid H_t=h,A_t=a_t)$, proving the result. The positivity condition ensures that this observed conditional distribution is defined. $\square$

This is a local identification step: it justifies using observed transition distributions to represent counterfactual transitions. Extending such steps through all decision times requires the full set of assumptions and an argument across the sequence; that is not accomplished by this one equality alone.

Worked Example: A Randomized Later Decision Given History

At a later decision, let $H_t$ indicate whether a patient’s current severity is high ($H_t=1$) or low ($H_t=0$). Suppose assignment is generated as $A_t=1\{U_t\leq q(H_t)\}$, where $q(0)=0.30$ and $q(1)=0.70$. Assume explicitly that $U_t\sim\operatorname{Uniform}(0,1)$ and is independent of the joint collection $(H_t,\mathcal Z_t)$, where $\mathcal Z_t$ contains the relevant potential future covariates and outcomes.

For either severity value $h$, independence gives the conditional distribution of $U_t$ given both $H_t=h$ and $\mathcal Z_t$ equal to its marginal distribution. Hence

P(A_t=1\mid H_t=h,\mathcal Z_t) =P(U_t\leq q(h)\mid H_t=h,\mathcal Z_t) =q(h) =P(A_t=1\mid H_t=h).

The treatment probabilities are therefore $0.30$ in the low-severity group and $0.70$ in the high-severity group, while assignment is conditionally unrelated to the potential future quantities. This satisfies the relevant sequential exchangeability condition. The treatment is not independent of severity, and it need not be: severity is part of the history being conditioned on.

The joint-independence assumption is important. It is not enough to say that $U_t$ is marginally independent of $\mathcal Z_t$. The next example shows why.

Why Marginal Independence Is Not Enough

Worked Example: Conditioning Can Create Dependence

Let $U$ and $Y^*$ be independent Bernoulli random variables, each with probability $1/2$ of equaling 1. Define a history indicator $H=1\{U=Y^*\}$ and treatment $A=U$. Marginally, $A$ and $Y^*$ are independent, so $P(A=1\mid Y^*=1)=P(A=1)=1/2$.

But among observations with $H=1$, the two variables must be equal. In particular,

P(A=1\mid H=1)=\frac12,\qquad P(A=1\mid H=1,Y^*=1)=1.

Thus $A$ and $Y^*$ are not independent conditional on $H$, despite their marginal independence. This constructed example is a warning about the logic of conditioning: a claim of marginal independence cannot establish sequential exchangeability given observed history. In an actual study, the treatment-generating mechanism must justify independence conditional on the history, not merely independence before conditioning.

From the Assumption to an Observed Transition

Worked Example: Interpreting a Conditional Transition

Suppose at a decision time the observed history $H_t$ includes a low-severity indicator and prior treatment. Among patients with one particular history $h$, 40 receive treatment $A_t=1$, and 24 of those have high severity at the next measurement. The observed transition estimate is

P(L_{t+1}=1\mid H_t=h,A_t=1)\ \text{estimated by}\ \frac{24}{40}=0.60.

If sequential exchangeability holds for $L_{t+1}^{\bar a_t}$, and the relevant treatment history has positive probability, the result above identifies this counterfactual probability with the corresponding population observed transition probability, estimated here by 24/40=0.60, subject to sampling uncertainty:

P(L_{t+1}^{\bar a_t}=1\mid H_t=h) = P(L_{t+1}=1\mid H_t=h,A_t=1).

The interpretation is limited and precise: it identifies a counterfactual next-covariate distribution for people with this observed history under the specified treatment through time $t$. It does not, by itself, identify the final outcome under an entire treatment strategy. That requires applying the corresponding conditions at later decisions as well.

How It Supports Regime Identification

A longitudinal treatment strategy specifies what treatment is to be given at each decision time, possibly as a function of the history. Sequential exchangeability supplies the no-unmeasured-confounding condition needed to use observed treatment-specific transitions as counterfactual transitions at each step. Consistency connects observed data to the treatment actually received. Positivity ensures the needed treatment choices are represented in the relevant histories. Together, these conditions permit the observed data to carry information about outcomes under strategies that were not necessarily followed by every individual.

The logic is sequential: condition on the history available at one decision, use exchangeability to remove selection on the counterfactual future at that decision, and use consistency to relate the observed transition to its counterfactual counterpart. The procedure must be valid at every decision along the strategy. A single randomized or unconfounded baseline assignment does not automatically guarantee the condition for later decisions.

Key distinction: Sequential exchangeability is an assumption about treatment assignment conditional on the observed history. It is not a property established merely by fitting a model, observing many covariates, or finding similar measured characteristics between treatment groups. Its credibility depends on whether the recorded history captures the factors that jointly predict later treatment and the relevant counterfactual future.

Assessment, Comparisons, and Failure Modes

Sequential exchangeability is generally not empirically testable from the observed data alone: the counterfactual outcomes under treatments not received are unobserved. Subject-matter knowledge and the timing and completeness of covariate measurement are therefore central. A useful assessment asks, at each decision, whether the history includes the clinical information that guided treatment and also predicts the future outcome. Covariates measured after a decision cannot serve as pre-decision confounders for that assignment.

The main alternative is not a different mathematical version of the same assumption, but a different design or analysis strategy: randomized treatment assignment can make exchangeability credible by design, while observational analyses rely on measured histories and substantive assumptions. Standard regression adjustment may condition on covariates, but under time-varying confounder feedback it does not automatically estimate the desired marginal regime contrast; the earlier material on that problem explains the concern. Marginal structural models use weighting to address this structure, but weighting does not repair unmeasured confounding or make a false exchangeability assumption true.

Omitted confounder: If an unrecorded symptom affects both $A_t$ and future outcome, conditioning on $H_t$ may not make treatment exchangeable.
Wrong timing: A variable recorded after $A_t$ is not part of the history available at that decision.
Overclaiming from baseline: Exchangeability at $t=0$ does not establish exchangeability at $t>0$.
Marginal-independence error: Independence before conditioning does not imply independence after conditioning on $H_t$, as the worked construction demonstrates.

Sequential exchangeability is a demanding but conceptually focused condition. It does not guarantee well-defined interventions, adequate support, or correct statistical estimation, and it cannot be verified by a diagnostic test. Its practical value is to make explicit the causal claim required at each treatment decision, so that the variables needed to defend that claim can be examined directly.

Check Your Understanding

Answer each question, then compare with the brief solutions.

Under the timing convention used here, what variables make up $H_t$, and why is $L_{t+1}$ excluded?
Does sequential exchangeability require $A_t$ to be marginally independent of $H_t$? Explain.
If $U_t$ is marginally independent of potential outcomes, is that sufficient to establish exchangeability conditional on $H_t$?
At one history $h$, the observed conditional transition probability is $0.25$. Under what assumptions can it represent a counterfactual transition probability?
Why does exchangeability at baseline alone not suffice for a multi-decision treatment strategy?

Solutions. (1) $H_t=(\bar L_t,\bar A_{t-1})$; $L_{t+1}$ is measured after $A_t$, so it is not available at that decision. (2) No. Treatment can depend on history; the assumption concerns its relation to relevant potential futures after conditioning on that history. (3) No. Conditioning can induce dependence; a sufficient randomization argument requires, for example, $U_t$ independent of the joint collection $(H_t,\mathcal Z_t)$. (4) It can represent $P(L_{t+1}^{\bar a_t}=l\mid H_t=h)$ when the relevant sequential exchangeability and consistency conditions hold and $P(A_t=a_t\mid H_t=h)>0$. (5) Later treatment choices can depend on evolving prognostic covariates, so exchangeability must be justified again at each later decision.