Emulating Staggered Treatment Initiation

Why Staggered Initiation Needs Calendar Alignment

In routine care, eligible patients rarely begin a treatment on one common date. Some initiate in January, others in March, and others later, as clinical practice, eligibility criteria, and background risks may change. A target trial emulation can represent this pattern as a set of trials indexed by calendar start date. The central design task is to ensure that treatment and comparator participants are eligible, classified, and followed from comparable dates.

Calendar time is not the same as follow-up time. Calendar time identifies when a person enters a trial; follow-up time measures how long it has been since that person's trial entry. If treatment initiators from a later period are compared with noninitiators from an earlier period, a difference in outcomes may reflect changing care or prognosis rather than treatment. Calendar alignment is therefore part of the design, not merely a modeling choice.

The protocol must specify which dates \(s\) are eligible trial starts: for example, the first day of each month or each quarter in a prespecified study period. For every date, it must apply the same clinical eligibility rule, covariate lookback window, treatment-strategy definition, outcome definition, and follow-up duration, except where a documented calendar change makes that impossible. A person can meet eligibility on more than one date; this is repeated eligibility, not a reason to silently treat the person as a different kind of participant.

Defining Eligibility, Assignment, and the Comparator

For a trial beginning at date \(s\), define \(E_{is}=1\) when person \(i\) is eligible and under observation on that date. A common new-user design requires that the person be alive, outcome-free for the outcome under study, have the required covariate history, and have no use of the treatment during a specified washout period ending at \(s\). The exact conditions depend on the clinical question; they must be applied identically to both strategies.

At \(s\), classify participants according to the protocol's assignment rule. For an immediate-initiation question, the treated group consists of eligible new initiators on \(s\), while the comparator group consists of eligible people assigned at \(s\) to a prespecified alternative, such as initiating an active comparator or not initiating the treatment. The comparator must be meaningful and defined at that same index date. It is not enough to select people who are untreated at some unrelated date.

If the target strategies allow initiation within a grace period rather than on the exact date \(s\), the protocol must say so explicitly and use the established strategy-assignment approach for that setting. In particular, it must not use subsequent initiation information to classify people at baseline without accounting for the resulting selection. Whatever the strategy, both groups begin follow-up at the same \(s\); follow-up duration is then expressed as time since that date.

Let \(L_{is}\) denote measured pre-assignment covariates for person \(i\) at trial start \(s\), including relevant calendar-time information if it varies within the chosen trial interval. The covariate window must close before assignment. For example, if the trial starts on April 1 and covariates are measured during the preceding 180 days, those measurements cannot include information from after April 1.

Calendar-Specific Effects and a Pooled Target

Let \(Y_{is}^{a}(\tau)\) be the potential outcome by \(\tau\) units of follow-up under strategy \(a\), for a person eligible at calendar date \(s\). For a binary outcome, define the date-specific risk among the eligible population as

A date-specific risk difference is \(\Delta(s,\tau)=R_1(s,\tau)-R_0(s,\tau)\). A pooled target across calendar starts should specify how much each start date contributes. If \(q_s\geq 0\) and \(\sum_s q_s=1\), the calendar-standardized risk difference is

For example, equal \(q_s\) gives equal weight to each trial start, while weights proportional to the number of eligible people at each start target a population emphasizing more populous starts. The weights define the target; they are not a technical detail that can be left implicit. If effects change over calendar time, \(\Delta_q(\tau)\) is a weighted average of those different effects, not a universal effect applying identically at every date.

To see the identification step, write \(Y(\tau)\) for the observed outcome under the strategy actually followed and \(A\) for baseline strategy assignment within trial \(s\). Conditional exchangeability gives \(Y^a(\tau)\) the same conditional distribution as \(Y(\tau)\) among people assigned \(A=a\), given \(L\). Consistency equates the observed outcome to the potential outcome for the strategy actually assigned. Therefore,

Positivity ensures the conditional mean on the right is estimable wherever the eligible population has covariate support. Subtract the identified risks for \(a=1\) and \(a=0\), then multiply each date-specific contrast by \(q_s\) and sum. Since the \(q_s\) are fixed and sum to one, the result is exactly \(\Delta_q(\tau)\). This argument identifies the target; it does not guarantee that an estimator will be precise or that the assumptions are credible. \(\square\)

A Baseline Assignment Rule Prevents Future-Based Classification

A central design check is whether a person's strategy can be determined using information available at the trial start. Consider the immediate-initiation protocol above. Both treatment and comparator participants must be eligible at \(s\), and each is classified at \(s\). Neither group earns entry by surviving untreated through a later period.

The justification follows directly from the assignment rule. Under baseline assignment, treated-group membership is determined by initiation at \(s\); it imposes no requirement about what happens during \((s,u)\) for any \(u>s\). Under future-based classification, membership requires observing a later initiation at \(u\), so a person who dies before \(u\) cannot enter that group. If the outcome includes death, or if survival is related to the outcome, the group definitions then encode a guaranteed survival interval for future initiators. Starting follow-up at \(s\) does not undo that selection. The correction is to define comparable strategy groups at the shared baseline, or to explicitly emulate a delayed-initiation strategy with an appropriate method for its grace period. \(\square\)

Practical Alignment Checks and Alternatives

Before estimation, inspect the design across calendar starts. Confirm that each trial has a defined eligibility date, that the same lookback and washout rules apply, and that treatment and comparator assignment use the same date and information window. Check that each participant's time zero is recorded consistently and that outcomes are assessed over equal elapsed follow-up, rather than merely to one common calendar end date. A common end date can give late entrants less follow-up and change the target if treated as though it were equal follow-up.

Also tabulate the number eligible and assigned to each strategy by trial start, and inspect the distribution of key baseline covariates by calendar period and strategy. Sudden changes may indicate altered coding, drug availability, eligibility, or comparator practice. A calendar-time effect model cannot repair a design in which one strategy is absent at a date or where its participants could not have been eligible under the same rules. Such a failure is a design limitation to report, not simply a reason to extrapolate.

There are several alternatives, each answering a different design question. A single trial beginning at one date avoids pooling across starts but may discard later eligible initiators and restrict generalizability. Separate date-specific trials preserve calendar-specific effects but may be imprecise. Pooling trials and standardizing to specified calendar weights produces one interpretable target, while regression adjustment for calendar time can be used to estimate date-specific risks or contrasts under an appropriate model. If treatment and comparator groups are matched within calendar risk sets, that can improve comparability, but matching does not itself define eligibility, establish a common time zero, or guarantee adequate overlap.

The limitation is substantive: a pooled estimate averages effects over the selected calendar period and weighting scheme. If clinical practice changes, the same nominal strategy may not represent the same intervention at every date. Changes in outcome recording or eligibility can also undermine comparability. State the calendar range, the trial-start schedule, and the target weights so readers can tell which population and era the estimate describes.

Common Pitfalls

• Using treated initiation dates but not comparator dates. This gives the two groups different time zeros. Assign comparators on the same prespecified calendar dates and begin follow-up at those dates.
• Defining treatment by a future start. This can make survival to that start a condition of treated-group membership. Classify strategies at baseline or explicitly handle the initiation window.
• Pooling crude risks with different calendar mixtures. The crude contrast can combine treatment effects with shifts in baseline risk. Use date-specific contrasts or standardize both strategies to the same \(q_s\).
• Assuming calendar adjustment solves noncomparability. Statistical adjustment cannot create a comparator group that was not available or make incompatible eligibility rules equivalent.
• Giving late entrants shorter follow-up without changing the target. Specify whether the outcome is assessed at a fixed elapsed time or at a common calendar end; these are different outcome definitions.