Emulating A Trial Of Acute Treatment Initiation

Why Acute-Care Emulations Need a Precise Clock

Acute treatments are often started within minutes or hours of a patient becoming eligible. That urgency makes the protocol’s timing consequential: a patient’s severity can change quickly, treatment decisions and outcome measurements may be recorded close together, and the data may contain orders without reliable administration times. A defensible emulation must say exactly when eligibility is assessed, when strategies begin, and how long outcomes are followed.

Consider a hypothetical emergency-department question about two clinically acceptable rescue medications for adults presenting with an acute migraine. The purpose here is to specify the trial that the observational analysis would emulate, not to recommend either medication. Treatment choice should be clinically plausible in both groups; if one option is contraindicated for a patient, that patient should not be treated as eligible for both strategies.

Specify the Acute-Care Trial Before Looking at Outcomes

Suppose the target trial enrolls adults at their first qualifying emergency-department encounter during the study period. Eligibility is assessed at the first documented clinical decision that rescue treatment is indicated, before either study medication is administered. A possible protocol requires a clinician-recorded acute migraine diagnosis, a baseline pain score in a prespecified range, no study medication already given during the encounter, and no contraindication to either option. It also requires enough prior records to measure prespecified baseline characteristics. These rules should be operationalized using fields and timestamps that actually exist in the data.

The strategies are to initiate medication A or medication B within 30 minutes of eligibility, using a prespecified dose and route. The first administration, not merely a medication order, defines initiation. Subsequent care is usual care under either strategy, including additional rescue medication if clinically needed. Thus, the contrast concerns the effect of the initial-treatment strategy, not the effect of taking only that medication throughout follow-up.

Time zero is the instant eligibility is first met and the treatment strategy would be assigned in the hypothetical trial. Follow-up starts at that same instant for everyone, including people whose medication is administered later in the 30-minute window. The primary outcome could be adequate symptom response by 2 hours after time zero, defined in advance as a pain-score decrease of at least 2 points from baseline. The outcome must be measured consistently enough in routine records to apply that definition. A separate protocol might use a clinically justified assessment window around 2 hours, but it must specify that window before examining results.

This estimand is an average effect in the eligible population. It is not automatically the effect among patients who actually received a medication, because actual treatment is selected in clinical practice. Nor is it the effect of treatment A versus no treatment: the comparator is initiation of B.

Work Through Eligibility and Time Zero

When strategy initiation is allowed during a grace period, patients cannot safely be classified at time zero simply by looking forward and placing them in the group matching the medication they eventually received. Doing so can create immortal-time bias: a patient must remain alive and observable long enough to receive the medication used to define the group. The protocol must handle the 30-minute period with an appropriate grace-period emulation, such as cloning compatible individuals to the strategies at time zero and censoring a clone when its observed care becomes inconsistent with that strategy, with suitable adjustment for informative censoring. The earlier material on treatment grace periods and cloning strategies develops those methods.

Assumptions Behind a Causal Interpretation

A carefully written protocol does not by itself make the comparison causal. The observational data must also support identification of the specified contrast. Let \(L\) denote measured characteristics recorded before time zero, including baseline pain, symptom duration, prior treatment, contraindications, comorbidities, and indicators of acuity that influence both treatment choice and response.

• Consistency: a person’s observed outcome under the strategy actually followed equals that person’s potential outcome under that strategy. This requires sufficiently well-defined versions of A and B, including dose, route, and what subsequent care the strategy permits.
• Conditional exchangeability: after accounting for \(L\), the treatment strategy followed is independent of the potential outcomes. In practice, this assumption is threatened if clinicians choose A or B based on unrecorded severity, contraindications, or patient preferences that also predict response.
• Positivity: for every relevant baseline history in the target population, there is a positive probability of following either strategy. If nearly all patients with a particular contraindication receive only one medication, they should not be included as though both strategies were feasible for them.
• Correct temporal and outcome measurement: eligibility, baseline covariates, administrations, and outcome assessments must be dated and recorded well enough to establish their order and apply the endpoint consistently. Loss of outcome observation must also be addressed under a prespecified rule.

Under these conditions, a baseline-adjusted comparison can identify the stated contrast. For example, if \(L\) is represented by prespecified strata and each stratum has support for both strategies, standardization expresses the target risk under strategy A as a weighted average of stratum-specific risks:

The weights are the distribution of \(L\) in the target eligible population. The analogous quantity for B uses \(P(Y=1\mid B,l)\), and their difference estimates \(\Delta\) if the assumptions hold. This is a justification for adjusting for baseline differences, not a guarantee that every relevant confounder has been measured. The treatment assignment model and adjustment method should follow the prespecified analysis strategy.

Check Feasibility and Diagnose Failure Modes

Before estimating an effect, check whether the protocol can be implemented in the actual records. Useful diagnostics include the proportion with a usable baseline pain score, the frequency and precision of administration timestamps, whether follow-up pain assessments are recorded near the planned time, and how often each treatment occurs among eligible clinical subgroups. These are not merely data-cleaning details: they test whether the emulation corresponds to the trial question.

Several patterns should prompt revision or caution:

• Unclear or delayed eligibility timestamps: if the time treatment was indicated cannot be distinguished from the time it was administered, baseline and exposure may be misordered.
• Orders substituted for administrations: an order may be canceled, delayed, or never given. Treating it as initiation misclassifies the strategy.
• Little treatment overlap: if one medication is virtually never used for patients with a given severity or contraindication profile, adjustment cannot reliably recover a comparison for those histories.
• Unequal outcome recording: if response measurements are more likely after one strategy, observed outcome differences may reflect measurement or follow-up processes.
• Changing treatment before initiation: if patients receive another rescue medication during the grace period, that event must be handled according to the prespecified strategy rules rather than ignored.

What This Acute-Treatment Contrast Does and Does Not Answer

An active-comparator trial of A versus B is often more clinically coherent than comparing A with no treatment when both are accepted rescue options. It can also make the eligible population clearer because both strategies must be feasible at the decision point. By contrast, an untreated group may differ sharply in indication, severity, or contraindications. An active comparator does not remove confounding, however; treatment selection can still depend on patient history and clinician judgment.

The 2-hour response outcome is suited to an acute question, but the endpoint must be clinically interpretable and reliably recorded. A different protocol could prioritize adverse events, need for additional rescue treatment, or return to care, but those endpoints have different measurement and follow-up requirements. Follow-up from time zero also means that delayed initiation leaves less time after the first dose for response to develop. That feature is part of the specified strategy contrast; changing the clock to start separately at each person’s administration would change the question and could introduce selection problems.

Finally, an initial-treatment strategy with usual subsequent care should not be interpreted as a comparison of sustained treatment regimens. Its result does not by itself describe the effect of continuing a medication over days or months. The protocol’s value is precisely its narrowness: it defines an acute decision, a short initiation window, a common outcome clock, and an explicit eligible population.