Why Time Zero Matters
An external control can meet the same eligibility criteria as a trial participant and still be observed from a different point in the course of illness. A trial may begin counting outcomes when a participant is enrolled, while external records may begin at diagnosis, a clinic visit, or treatment initiation. These dates are not interchangeable. They determine which information is considered baseline and which events fall inside follow-up.
For an external-control comparison, time zero is the common conceptual start of eligibility, treatment assignment or decision, and outcome follow-up. Each external patient needs an index date that represents that point as closely as the data allow. The purpose is not to make calendar dates identical across patients. It is to make the clinical position from which follow-up begins comparable.
The trial protocol determines the reference point. If trial follow-up starts at randomization or enrollment, an external index date should represent the corresponding eligibility and treatment-decision point—not automatically the external patient’s first diagnosis or first recorded treatment. If trial follow-up starts at first dose, the protocol must instead define a suitable external counterpart for that origin. The comparison should not use one origin for trial participants and another for controls simply because one is easier to find.
This alignment is necessary for interpreting baseline characteristics and outcome risks. It does not, by itself, make the groups exchangeable or remove confounding. It also does not settle how routine-care treatment strategies are defined; those require a separate design decision.
Operationalizing an External Index Date
A defensible date rule starts with a clinical and temporal definition, then identifies which data fields can implement it. For example, the rule might use the first documented visit on which a patient satisfies the trial’s eligibility criteria and a treatment decision is made. The chosen field should correspond to that event, rather than merely being a date that is consistently available.
The rule should be specified before comparing outcomes. It should say which event anchors the date, how ties or multiple qualifying visits are handled, and what happens when the relevant date is missing or too imprecise. Eligibility must be established using information available at or before the index date. A date chosen because the patient later received a particular treatment, survived long enough to be treated, or had a later outcome uses future information to define entry and can select a different population.
Identify the protocol-defined date at which trial patients enter the comparison and follow-up begins.
Describe the clinical event that represents the same point for an external patient, using information available by that date.
State how records identify the event, handle repeated qualifying events, and address imprecise or unavailable dates.
Check that eligibility, baseline measurements, and the beginning of outcome ascertainment are anchored to the intended time zero in both groups.
A useful formal check is to define the risk being compared. Let \(T\) denote time to the outcome and \(t_0\) the aligned origin. For a fixed follow-up duration \(\tau\), the event risk is
The condition that patients are outcome-free at \(t_0\) applies when the endpoint and estimand require it; it is not a universal requirement for every outcome. The key is that the origin, eligibility condition, and follow-up duration refer to the same point in both groups. If one group is counted from a different \(t_0\), its numerator and risk set may describe a different question.
Worked Examples
Worked Example: An Earlier Diagnosis Date Creates a Longer Window
Suppose a hypothetical trial follows 100 participants for 90 days from enrollment. Ten participants experience at least one nonfatal outcome during those 90 days, so the trial’s observed 90-day risk is \(10/100=0.10\), or 10%. An external data set contains 100 eligible patients. Investigators can identify a qualifying clinical decision date, but initially count external outcomes from a diagnosis date 60 days earlier.
Between the earlier diagnosis date and the aligned index date, 6 external patients experience the outcome. During the 90 days after the aligned index date, 10 patients experience it. These two sets overlap: 2 of the 6 patients with an earlier outcome also have an outcome in the post-index window. Therefore the number of unique patients with an outcome anywhere in the combined 150-day period is \(6+10-2=14\), not 16. The combined-window patient risk is \(14/100=0.14\), or 14%, for that 150-day window.
That 14% is not an estimate of the external 90-day risk comparable to the trial’s 10%. Restricting the external count to the aligned 90 days gives \(10/100=0.10\), or 10% in this hypothetical example. The agreement is illustrative, not evidence that the groups are otherwise comparable. This calculation also relies on patient-level identification of unique outcomes; simply adding counts from overlapping periods would double-count some patients.
Worked Example: Baseline Disease Duration Depends on the Index Date
In a hypothetical trial, patients enter at a documented treatment-decision visit. Their median disease duration at enrollment is 3 months. An external cohort has records of diagnosis and clinic visits. If the index date is incorrectly set to diagnosis, the external baseline disease duration is 0 months by construction. If the intended external counterpart is the treatment-decision visit, the median duration by that visit is 3 months.
The calculation is direct: for each patient, disease duration at baseline is the index date minus the diagnosis date. Under the diagnosis-based rule, this difference is zero. Under the decision-visit rule, the cohort median difference is 3 months. Thus, a comparison using diagnosis as external time zero would compare trial patients enrolled after a period of disease with external patients observed at diagnosis. A baseline table could show this discrepancy, but the more important point is that the two groups would enter the comparison at different clinical stages.
Using the decision visit as the external index date is appropriate only if it genuinely represents the trial’s entry point and is recorded consistently. A date that happens to produce a similar median disease duration is not sufficient; the event represented by the date must also be clinically aligned.
Worked Example: Quantifying the Consequence of Shifting Follow-Up
Suppose a hypothetical external cohort has 80 eligible patients. From the aligned index date, 8 unique patients have an outcome within 90 days. Its aligned-window risk is
Now suppose the analyst instead uses a date 30 days earlier and counts outcomes over 120 days. In that earlier 30-day period, 5 patients have an outcome; all 5 are distinct from the 8 patients with outcomes during the aligned 90 days. There are therefore \(5+8=13\) unique patients with an outcome over the full 120-day window, giving \(13/80=0.1625\), or 16.25%.
The 16.25% figure is a valid patient risk for the specified 120-day window under these assumptions, but it is not the aligned 90-day risk. The assumption that the two sets are distinct is stated explicitly here; if patients could appear in both sets, the number of unique patients would instead be the size of their union. Changing the index date changed the question being answered, not merely the way the same risk was calculated.
Derivation: Why a Shifted Origin Changes Risk
Let the intended origin be \(t_0\), and consider a shifted origin \(t_0-d\), where \(d>0\). A risk over duration \(\tau\) from the intended origin concerns events between \(t_0\) and \(t_0+\tau\), among patients eligible at \(t_0\). A risk over duration \(d+\tau\) from the earlier origin concerns events between \(t_0-d\) and \(t_0+\tau\), among patients eligible at the earlier date. These intervals differ, and so may their eligible populations.
Even if the same patients could be followed across both intervals, the earlier-window outcome indicator is not generally the same as the aligned-window indicator. For a patient-level “any outcome” endpoint, the earlier window’s indicator is 1 if the patient has an outcome anywhere in the longer interval. It can be written as the maximum of the indicators for the pre-index and post-index subperiods. Thus, its cohort risk depends on both subperiods and their overlap. It cannot be interpreted as the post-index risk alone.
This is why a matching follow-up duration is not enough if the origins differ, and why matching a calendar date alone is not enough if patients are at different points in their clinical course. The protocol-defined origin and the external operational rule must work together.
Diagnostics, Alternatives, and Limitations
Several checks can reveal that the date rule is not working as intended:
- Baseline timing: Compare time since diagnosis and other time-sensitive characteristics at the index date. Large discrepancies can indicate that the cohorts enter at different stages, though similar summaries do not prove alignment.
- Event ordering: Verify that eligibility information and baseline covariates precede or coincide with the index date, and that counted outcomes occur after it under the chosen definition.
- Date distribution: Inspect whether dates cluster at data-system changes, annual visits, or treatment milestones rather than at the intended clinical event.
- Record completeness: Check the proportion with precise, usable index dates and whether missingness differs across clinically important subgroups.
- Sensitivity to plausible rules: Recalculate descriptive baseline and outcome summaries under defensible alternative date rules. Material changes indicate that conclusions depend on the operationalization and should be reported as such.
The main alternatives are not equally defensible. A diagnosis date is easy to find but may precede trial entry by months or years. A first-treatment date may be suitable only when it corresponds to the trial’s origin; for untreated external controls, an analogous decision point must be specified rather than assumed. A fixed calendar date is useful for defining a period but does not alone establish that patients were at the same clinical stage. The preferred rule is the one that represents the trial’s origin, is observable in the external data, and does not rely on future information.
Some records may not contain a date precise enough to establish the intended point. Others may only show that a patient met criteria sometime within a broad interval. In such cases, the index date is uncertain, and the analyst should not present it as exact. Depending on the data and design, a restricted cohort or sensitivity analysis may be more credible than assigning a convenient but clinically unrelated date.
Common Mistake
Wrong reasoning: “The diagnosis date is available for nearly everyone, so it is a good baseline date; we can simply compare the outcome percentages.”
Correction: Availability does not establish clinical alignment. First identify the trial’s time zero, then determine whether the diagnosis date represents that same point for external patients. If it does not, use a defensible counterpart to the trial origin or clearly acknowledge that the cohorts and outcome windows differ. Also count unique patients for patient-risk calculations: counts from two periods can be added only when the groups are known to be disjoint; otherwise, use the number of unique patients in their union.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A trial starts follow-up at enrollment. An external cohort is indexed at diagnosis, typically 5 months earlier. Name two ways this can undermine comparison.
- In a hypothetical external cohort of 50 patients, 4 have an outcome before the aligned index date and 7 have an outcome after it during the intended follow-up. Two patients are in both sets. How many unique patients have an outcome across both periods, and what does the resulting proportion represent?
- Why is choosing an external index date based on a treatment received several weeks later potentially problematic?
- Trial follow-up starts at first dose. What should be done before assigning external patients an index date?
- Does aligned time zero guarantee that the external and trial groups are exchangeable? Explain.
Solutions. (1) External patients may enter at an earlier disease stage, and their outcomes may be counted over a different interval; baseline characteristics can consequently refer to different clinical moments. (2) The unique count is \(4+7-2=9\), so the proportion is \(9/50=0.18\), or 18%, for the combined period covered by both sets. It is not automatically the risk in the intended post-index follow-up, which is \(7/50=14\%\) if the 7 are unique patients in that window. (3) It uses future information to define entry and can select patients based on what happens after the proposed origin. (4) Specify what external clinical event represents the first-dose origin, and verify that it can be identified without relying on future outcomes or treatment information in a way that changes eligibility. (5) No. Alignment makes the time origin more coherent but does not remove confounding or guarantee that observed groups are otherwise comparable.