Why a Grace Period Needs a Precise Rule
A patient may be eligible for a treatment strategy today but start treatment several days later because of scheduling, clinical assessment, or patient preference. A grace period makes that delay part of the strategy being studied. It is not merely a convenient allowance in the data: its length, starting point, and rules for what happens before initiation help define the causal question.
The key design requirement is to define the strategy at eligibility and time zero, before observing who actually starts treatment. Otherwise, people can be assigned to a group based on what happens to them later. In particular, a person whose endpoint occurs before treatment can no longer start afterward. Excluding such a person from the group defined by future treatment initiation can make the groups incomparable from the outset.
Write \(S\) for the observed time of treatment initiation, with \(S=\infty\) if treatment is never initiated during the relevant follow-up. Write \(T_Y\) for the time of the endpoint. These observed times describe what happened; they do not, by themselves, define treatment assignment at time zero. For a causal question, the strategy must state what treatment course would be followed under the assigned strategy, including the rule for starting within the window.
For example, “initiate by day 14” defines a deadline but may not fully specify the strategy if there are many clinically acceptable start times. One protocol might start as soon as a prespecified assessment is complete; another might allow the clinician and patient to choose any start day through day 14. If the timing choice can affect the outcome, the protocol needs to make that choice rule explicit or recognize that it is comparing a set of possible treatment courses rather than one fully specified course.
Different Windows Mean Different Strategies
A strategy to initiate by day 7 is not the same strategy as one to initiate by day 30. Nor is either necessarily equivalent to “start immediately.” If early treatment has effects, changing the deadline changes the treatment courses allowed by the strategy and can change the causal estimand. Both groups should share the same eligibility criteria, time zero, outcome definition, and follow-up origin if the intended contrast is between initiation policies.
| Strategy description | What it permits or requires | Design question to settle |
|---|---|---|
| Initiate at time zero | Start at eligibility, subject to an explicit feasibility rule | What counts as initiation at time zero? |
| Initiate within 7 days | Start during \([0,7]\) | What start-time rule applies within the window? |
| Initiate within 30 days | Start during \([0,30]\) | Are all permitted start times considered part of this policy? |
| Do not initiate during the specified period | Remain untreated over the stated period | What treatment is allowed after that period? |
A useful target-trial description therefore says more than “treated versus untreated.” It identifies the deadline, the treatment and dose or version where relevant, the start-time rule, and what happens to people who have not started by the deadline. A grace period also does not automatically define the treatment strategy after initiation: continuation, discontinuation, or switching may require further rules.
Worked Example: Two Different Initiation Deadlines
Suppose eligible patients enter follow-up at time zero. The study compares a policy to start Drug A by day 7 with a policy to start Drug A by day 30. Assume, for this example, that each policy says to start on the first day the patient completes a specified clinical assessment, provided that the patient is still alive and has not had the endpoint. The assessment occurs on day 5 for one patient and day 12 for another.
Day-5 patient: the patient can comply with either policy by starting on day 5. The start is within both \([0,7]\) and \([0,30]\).
Day-12 patient: the patient can comply with the 30-day policy by starting on day 12, but cannot comply with the 7-day policy, whose deadline has passed. That difference is not a data-labeling technicality: the strategies allow different treatment histories.
If an endpoint occurs on day 4, before either patient has started, the endpoint is part of follow-up under the assigned strategy; the patient cannot be required to start on day 5 or day 12 afterward. The protocol must count the endpoint rather than silently remove the patient from the analysis because no later initiation was observed.
Endpoints Before Initiation
The strategy must account for events that occur during the grace period. Suppose treatment can be initiated only while a person is alive and endpoint-free. If an endpoint occurs before any initiation, that person cannot subsequently satisfy a rule requiring initiation within the window. But it is incorrect to conclude that all endpoints by day \(g\) are absent among people who initiated by day \(g\): a person can initiate first and then experience the endpoint before the deadline.
Proof. Under the stated assumption, a person with no initiation before \(T_Y\) cannot initiate at or after \(T_Y\). Thus no initiation occurs at any time, so \(S=\infty\), which implies \(S\not\leq g\). However, if \(S<T_Y\leq g\), both \(S\leq g\) and \(T_Y\leq g\) hold. The endpoint prevents initiation afterward, not an endpoint after initiation. Therefore the implication applies only to an endpoint before any initiation, not to every endpoint within the window. \(\square\)
Worked Example: Ordering Initiation and an Endpoint
Consider three people, with a 10-day grace period. Initiation is possible only before an endpoint. Person A initiates on day 3 and has the endpoint on day 8. Person B has the endpoint on day 4 without having initiated. Person C initiates on day 12 and has no endpoint by day 10.
For Person A, \(S=3\leq10\) and \(T_Y=8\leq10\). This person both initiated within the grace period and experienced the endpoint during it.
For Person B, the endpoint occurs before any initiation. No later initiation is possible, so \(S=\infty\), and this person does not meet the observed condition \(S\leq10\). The endpoint still occurred during follow-up; it must not be erased by classifying the person according to later treatment.
For Person C, \(S=12>10\), so the person did not initiate within the window. The example illustrates why a strategy needs an explicit rule for a missed deadline and why the observed groups “initiated by day 10” and “not initiated by day 10” are not automatically valid groups for causal comparison.
The Risk of Assigning Exposure From Future Treatment
A tempting analysis labels everyone who eventually initiates within \(g\) as “treated” from time zero, and everyone else as “untreated.” That label uses post-baseline information. To enter the first group, a person must remain alive and endpoint-free long enough to initiate if the endpoint prevents later initiation. This creates a selected group with a guaranteed event-free interval before its observed initiation time. It does not guarantee that the person remains event-free through all of \([0,g]\): an endpoint after an early initiation can still occur within the window.
The distinction can be written directly. Under the assumptions of the result above, the event \(\{S\leq g,\ T_Y<S\}\) is impossible: an endpoint before initiation prevents a later start. But \(\{S\leq g,\ S<T_Y\leq g\}\) need not be impossible. Thus, future-initiation classification can omit early endpoints from the eventual-initiator group while retaining endpoints that occur after initiation. This asymmetry can distort a comparison, though the direction and size of the resulting bias are not determined by the window alone.
Worked Example: How Future Classification Changes Group Membership
Imagine 100 eligible people and a 7-day grace period. Assume each endpoint prevents later initiation. By day 7, 20 people have initiated: 10 of those experience the endpoint after starting, and 10 do not. Another 10 people experience the endpoint on day 2 before starting and therefore never initiate. The remaining 70 people neither initiate by day 7 nor have the endpoint by then.
If “treated” is defined as \(S\leq7\), the treated group has 20 people, of whom 10 have the endpoint by day 7: its observed proportion is \(10/20=0.50\). The group includes post-initiation endpoints.
The “not treated by day 7” group has 80 people: the 10 early endpoints plus the 70 people without an endpoint by day 7. Its observed proportion is \(10/80=0.125\). The arithmetic is correct, but these are not groups assigned at time zero. Membership in the eventual-initiator group required having an opportunity to initiate before an endpoint. The contrast \(0.50-0.125=0.375\) is therefore not, on its own, a causal effect of the two initiation policies.
This example does not establish that future classification always biases an estimate in one direction. Different treatment effects, baseline risks, start-time patterns, and other causes of initiation can change the observed contrast. It establishes the more basic design problem: the group label depends on post-baseline events and is not equivalent to assignment to a strategy at time zero.
How to Specify and Check a Grace-Period Design
The following sequence makes the design choices visible before analysis. It does not itself prescribe a particular estimator; the target-trial protocol must first define the contrast that an estimator would target.
Eligibility, assignment to a strategy, and the start of outcome follow-up must be aligned. Define the grace-period clock from this point.
State the deadline, the allowed treatment, the timing rule within the window, and the rule for treatment after the deadline.
Say whether initiation is no longer possible, how the event contributes to the outcome, and how any other event that prevents initiation is handled.
Verify that eligibility, treatment initiation, endpoints, and the deadline are recorded on a compatible time scale. Decide how same-day ordering and missing initiation dates are handled.
Diagnostics should look for whether the observed data can represent the stated rules: Are start dates available with enough precision to distinguish initiation before an endpoint? Are there people who start after the stated deadline? Are endpoints recorded during the window, including among people who later initiate? Do treatment versions or timing choices differ in ways the strategy did not define? These checks can reveal a mismatch between the protocol and the available data. They cannot establish that treatment groups are exchangeable or that a causal estimate is unbiased.
A common pitfall is to call people “treated” only after seeing that they eventually started, then count their follow-up from time zero as treated follow-up. The incorrect reasoning is that eventual initiation proves they were treated “for the strategy” at baseline. It does not: before their start, they had not yet received treatment, and they had to remain eligible to reach that start. The correction is to define the strategy at time zero and preserve outcomes that occur before initiation. An analysis that emulates such strategies must account for the timing and rules explicitly; subsequent material develops particular approaches for doing so.
How the Main Alternatives Differ
An immediate-start strategy avoids a grace period but may be clinically unrealistic or unsupported by the data. A fixed baseline-assignment strategy instead compares assignment to policies at time zero, regardless of later adherence; that is a different question from the effect of following a start-by-\(g\) rule. A grace-period strategy allows delayed initiation but requires explicit timing and event rules. Simply comparing observed initiators with observed noninitiators does not substitute for any of these assigned strategies.
The earlier material on intention-to-treat analogs and per-protocol effects distinguishes assignment to a strategy from following that strategy. The same distinction matters here: “assigned to initiate by day 30” is not synonymous with “observed to initiate by day 30.” A grace period makes treatment timing part of the policy; it does not turn future behavior into baseline assignment.
The limitation is substantive as well as statistical. A long window may combine meaningfully different treatment timings, while a short window may be infeasible for many eligible patients. If the timing rule is vague, the strategy and its causal interpretation are vague too. State the policy at the level the clinical question requires, and do not infer a universal best grace-period length from the definition alone.
Check Your Understanding
Answer each question, then compare with the brief solutions below.
- With time zero at eligibility, what does a 14-day grace period specify, and what additional detail may be needed?
- Can someone both initiate within a 10-day grace period and experience the endpoint by day 10? State the required ordering.
- If an endpoint occurs on day 3 before any initiation, and the endpoint prevents later initiation, can that person subsequently satisfy \(S\leq7\)? Explain.
- Why is classifying people as treated from time zero based on observed initiation by day 7 problematic?
- Give one reason that “initiate by day 7” and “initiate by day 30” define different strategies.
Solutions. (1) It specifies an initiation window \([0,14]\); the protocol may also need a rule for when within the window to start and what happens if initiation does not occur. (2) Yes. If initiation occurs first, for example \(S=4\) and \(T_Y=8\), then both times are within 10 days. (3) No. The endpoint occurred before any initiation and prevents a later start, so \(S=\infty\) under the stated convention. (4) Group membership uses future information and, when endpoints prevent later initiation, excludes some early endpoints from the eventual-initiator group. (5) The policies allow different start times: a start on day 20 meets the 30-day deadline but not the 7-day deadline.