Is sequential exchangeability the same as positivity?

No. Positivity concerns available treatment alternatives; exchangeability concerns confounding conditional on history.

Does randomizing the first treatment settle every later comparison?

No. Later treatment receipt or decisions may introduce additional selection.

Can the assumption be fully tested from observed outcomes?

Generally no. It involves potential outcomes that are not jointly observed, so substantive knowledge and sensitivity analysis matter.

AI Glossary / Causal inference and experimentation / Advanced

Sequential exchangeability

Definition

Sequential exchangeability is a causal assumption that, at each treatment decision, treatment is independent of the relevant future potential outcomes conditional on the observed history available before that decision. Informally, the history is sufficient to address treatment-outcome confounding at every stage, not merely at the start of follow-up.

What it means

Repeated decisions respond to changing circumstances. A manager may assign training this month because workload or performance changed last month. If those circumstances also influence the final outcome, a causal analysis needs to account for their role at the appropriate time. Sequential exchangeability says that, after conditioning on the specified history, no additional information driving treatment selection also predicts the relevant potential outcomes in a way that confounds the comparison. It is a substantive assumption about the process, not a property that software can certify from a fitted model.

How it works

Define the treatment sequence, potential outcomes and history at each decision point. Examine why each treatment choice occurs and whether the measured history captures the required common causes. Together with consistency and positivity, suitable sequential exchangeability conditions support longitudinal identification methods such as the g-formula and inverse probability weighting. Baseline randomization alone does not ensure exchangeability for later self-selected treatment decisions. The assumption is separate from correct model specification: a perfectly fitted model of an insufficient history can still leave confounding unresolved.

Three practical examples

A coaching study records workload before each monthly coaching decision because workload influences both the decision and later performance.
A maintenance analysis records equipment condition before each service action. If an engineer's unrecorded assessment also drives service and predicts failure, the measured history may be insufficient.
A sequential experiment randomizes assistance at each eligible decision point using known probabilities. The design supports the relevant assignment comparison, provided the analysis respects eligibility, timing and the intervention being studied.

Think of it this way

Think of pausing a film before every decision and asking whether the information on screen explains the reasons for that choice that also matter for the later outcome. If an important off-screen factor drives both, the comparison remains vulnerable. Checking the opening scene is not enough when later decisions respond to new developments.

Worked example

Consider an original hypothetical second-month coaching decision among staff with the same recorded history. Suppose an unrecorded motivation factor affects future performance. Highly motivated staff receive coaching with probability 0.8, while less motivated staff receive it with probability 0.2. Even though both treatment options occur, treatment selection still carries information about a predictor of potential outcomes within the recorded-history group. Sequential exchangeability with respect to that history therefore fails in the stipulated example. The nonzero probabilities illustrate that positivity can hold while exchangeability fails. Measuring a useful motivation proxy might help, but a proxy does not automatically capture the confounding factor sufficiently.

Illustrative case study

A fictional operations company evaluates a sequence of optional coaching sessions. Its records contain baseline role and experience, but supervisors later assign sessions based on changing case difficulty. Analysts discover that case difficulty is available in weekly logs and align it to the date before each coaching decision. They also learn that supervisors use informal assessments not captured in those logs. The report explains why the added measurements improve the recorded history without proving that it is sufficient. The team considers sensitivity analysis and a future sequentially randomized design. This case illustrates assumption assessment, not a demonstrated causal effect.

Common misconceptions and mistakes

Sequential exchangeability does not require treatment to be marginally random or unrelated to history. Treatment can depend strongly on measured history. The requirement concerns the remaining relationship with potential outcomes after conditioning. Another error is using a measurement taken after a decision as though it were available beforehand. Analysts can also mistake good covariate balance or strong treatment prediction for proof of no unmeasured confounding. Neither establishes the assumption. More rows and more flexible algorithms do not automatically reveal an omitted cause of treatment and outcome.

Business use

The assumption helps organizations understand which operational records are necessary for evaluating adaptive programs. It encourages teams to document why decisions are made and what information was available at the time. This can reveal gaps that a final performance dashboard misses. Business reports should explain important potential confounders and the limits of their measurement in plain language. Where key decision drivers are unrecorded, the analysis may support a qualified conclusion or motivate better data collection and study design rather than a confident causal claim.

Check your understanding

Why did exchangeability fail in the worked example?

Answer: Unrecorded motivation affected both treatment selection and potential outcomes within the recorded-history group.

Why did positivity still hold?

Answer: Both treatment probabilities were between zero and one.

What is the first timing check for a proposed adjustment variable?

Answer: Whether it was measured before the treatment decision it is meant to help explain.

Related terms

Conditional exchangeability
sequential randomization
time-varying confounding
positivity
g-formula
longitudinal causal inference

Learn next

Build a two-stage decision timeline and list the information used at each stage. Mark unrecorded factors that could also influence the outcome. Then compare observational adjustment with a sequentially randomized design for the same intervention question.

Sources

Causal Learning Handbook, Longitudinal Causal Inference and Time-Dependent Confounding: https://amertens.github.io/causal-learning-handbook/03-07-longitudinal-td-confounding.html Hernán and Robins, Causal Inference: What If, longitudinal treatment chapters: https://www.hsph.harvard.edu/miguel-hernan/wp-content/uploads/sites/1268/2024/04/hernanrobins_WhatIf_26apr24.pdf The motivation probabilities and coaching scenario are original educational examples with stipulated causal relationships.

Previous: Sequential decision making Next: Sequential Feature Selection