Risk-Set Matching: Why the Comparison Group Must Follow the Clock
Anas H. Alzahrani, MD PhD MPH
Department of Preventive Medicine and Public Health
Faculty of Medicine, King Abdulaziz University
Risk-set matching solves a deceptively simple problem: when treatment can begin at any moment, who counts as untreated? In an emergency-care registry, one patient may receive defibrillation at minute three, another at minute seven, and a third never before the ambulance arrives. Sorting them into fixed “treated” and “untreated” groups uses tomorrow's information to rewrite who was comparable today.
The cleaner design freezes the clock whenever treatment starts. It compares the newly treated patient with people who were still eligible, still under observation, and not yet treated at that same moment. The untreated group is not a permanent waiting room. It is a sequence of risk sets.
The Bias Begins When Exposure Is Defined Using the Future
Imagine classifying everyone after the resuscitation ends. Anyone who ever received public-access defibrillation goes into the treated group; everyone else becomes a control. That looks reasonable until you ask what had to happen for someone to be treated at minute ten. The patient had to remain in cardiac arrest, remain in a setting where defibrillation was possible, and avoid earlier competing events such as return of spontaneous circulation or emergency-service arrival.
Those conditions are prognostic. Longer resuscitation both creates more opportunities for later treatment and marks a changing clinical course. A fixed exposure label therefore mixes treatment with the time-dependent process that made treatment possible. Baseline propensity-score matching cannot repair a comparison set assembled with future information.
The clock rule
At the moment a patient starts treatment, compare them only with patients who could also have started treatment at that moment. Do not let later treatment or later outcomes decide earlier eligibility.
Interactive risk-set explorer
Freeze the clock before choosing controls
Pick a treatment minute. A valid comparator must still be eligible, untreated, and under observation at that exact moment.
Index treatment
Patient A receives public-access defibrillation at minute 3.
2 of 4 candidates belong in this moment's comparison pool.
Patient B
Eligible nowNot yet treated; receives PAD at minute 7
Future treatment does not erase eligibility at minute 3.
Patient C
Eligible nowNever receives PAD before EMS arrival
Still under observation and eligible at minute 3.
Patient D
Exclude nowReceived PAD at minute 2
Already treated before this risk set was formed.
Patient E
Exclude nowAchieved return of spontaneous circulation at minute 2
No longer at risk of receiving the intervention.
Teaching simplification: real analyses also match on measured history available by each minute and must account for repeated use of comparators, matching choices, and within-pair dependence.
What Sequential Risk-Set Matching Actually Does
- Define the changing eligibility state. Specify what makes a patient able to receive treatment at each time point and which events remove that possibility.
- Open a risk set when treatment begins. Each newly treated patient creates a local comparison problem at their treatment time.
- Use only history available by then. Estimate treatment propensity from baseline and time-updated variables measured no later than that moment.
- Match within the moment. Select one or more untreated, eligible comparators with similar measured history from that time-specific pool.
- Analyze the matched contrasts correctly. Respect matched sets, repeated comparator use if allowed, and the outcome time scale defined by the estimand.
The counterintuitive move is step two: a patient treated later may serve as a comparator earlier. That is not contamination. It is faithful to what was known then. Excluding that patient because of future treatment would condition the earlier comparison on an event that had not happened yet.
Fixed Matching and Risk-Set Matching Answer Different Design Problems
| Design question | Fixed baseline groups | Sequential risk sets |
|---|---|---|
| When is treatment assigned? | At one shared baseline | At many possible times during follow-up |
| Who can be a comparator? | Eligible and untreated at baseline | Eligible and untreated at the index patient's treatment time |
| Can a later-treated patient be a control? | Usually classified as treated for the whole study | Yes, before their own treatment begins |
| Main failure if misused | Future exposure edits the earlier comparison | Sparse late risk sets or poorly measured time-updated confounding |
A Clinical Example: Defibrillation While the Clock Is Moving
A 2026 nationwide Japanese cohort studied public-access defibrillation among people with bystander-witnessed ventricular fibrillation in public locations. The investigators sequentially matched a patient receiving defibrillation with a patient still at risk of receiving it in the same minute. After matching, the analysis included 1,664 pairs and reported an association with better one-month neurological outcome: risk ratio 1.46, with a 95% confidence interval from 1.34 to 1.59.
The method is the important lesson here, not a claim that matching converted the registry into a randomized trial. The study still relies on measured confounding control, correct time recording, appropriate eligibility rules, and an interpretable treatment contrast. The estimates for later timing windows were much less precise, which is expected as patients and valid comparators disappear from the risk set. A dramatic late point estimate does not compensate for a very wide interval.
Five Questions Reviewers Should Ask
1. What exactly is the time-varying estimand?
“Treat now versus do not treat now” is not necessarily “ever treat versus never treat.” If comparators may receive treatment later, say whether the contrast represents initiation now, a delay, or a treatment-policy strategy.
2. Who enters and leaves each risk set?
Eligibility, treatment, outcome, recovery, competing events, and administrative transitions should be timestamped. A vague “untreated at that time” label is not enough to reconstruct the design.
3. Were covariates measured before the matching decision?
A variable recorded after treatment initiation cannot justify the match. Time-updated severity may be necessary, but its measurement cadence and causal role need explicit defense.
4. Does overlap survive late in follow-up?
Report the number treated, number eligible, successful matches, covariate balance, and uncertainty by time window. Late estimates often describe a small, selected population.
5. How were reuse and dependence handled?
A comparator may appear in several risk sets or later become treated. The analysis and standard errors must reflect the actual matching and reuse structure rather than pretending every row is independent.
When Risk-Set Matching Is Not Enough
Matching balances measured history under the chosen model. It does not recover an unrecorded shockable rhythm, bystander skill, device accessibility, evolving severity, or clinical judgment that affects both treatment and outcome. Nor does good overall balance establish positivity in every minute. Time-dependent treatment can also create post-treatment confounding problems that require other designs, including cloning with censoring and weighting or marginal structural models.
The decision rule is practical: use sequential risk sets when treatment initiation is staggered and the comparison must be reconstructed at each treatment time. Then audit the estimand, timestamps, time-updated confounders, overlap, and analysis structure. A method name is evidence that the authors noticed the clock. It is not evidence that every clock-related bias is gone.
Why This Matters for Aqrab
Timing errors often hide inside competent-looking methods sections. Aqrab is built to ask whether eligibility, treatment assignment, comparison selection, covariate history, and follow-up align with the claim—not merely whether “propensity score matching” appears in the text.
Paste a protocol or observational methods section into Aqrab Try to pressure-test the time axis and comparison logic. For repeatable registry or manuscript-review workflows, explore the developer tools.
Methods Anchors
The clinical example follows the 2026 public-access defibrillation cohort. The risk-set matching logic follows Lu's time-dependent propensity-score framework and the later review of matching with time-dependent treatments. These methods can improve design alignment under measured-confounding assumptions; they do not establish exchangeability from observed balance alone.
Keep reading
Don't stop at one method.
Good methods judgment comes from contrast. Read the neighboring guides, see where the assumptions diverge, and avoid treating every observational problem like it needs the same hammer.
Time-Varying Treatment Dose: When Recovery Gets Counted as the Intervention
A practical guide to time-varying treatment dose in clinical research. Learn why accumulated exposure can encode recovery, how baseline matching misses treatment-confounder feedback, and what reviewers should demand from longitudinal dose-response claims.
Causal Readiness: When a Huge Linked Dataset Still Cannot Identify an Effect
A practical guide to causal readiness in linked health and administrative data. Learn why scale and propensity-score overlap are not enough when treatment, need, comparators, or outcomes are poorly measured.
Weak Instruments and Physician Preference IVs: When Treatment Movement Is Not Yet Causal Credibility
A practical guide to weak instruments and physician-preference IVs for clinical researchers. Covers first-stage weakness, exclusion leakage, local interpretation, and what reviewers should demand before trusting an IV claim.