Calendar-Time Trends in Administrative Censoring

Why a Common Study End Can Create a Calendar-Time Problem

When everyone is observed until a shared study end, participants who enter earlier can contribute more follow-up than those who enter later. That unequal follow-up is the ordinary consequence of administrative censoring. The harder question is whether the cohorts entering at different calendar dates have comparable event processes. If event rates, clinical care, or eligibility rules change over calendar time, they may not.

This distinction matters because a routine analysis indexed only by time since entry can combine different calendar-time processes. At a given follow-up duration, the people still observed may disproportionately come from earlier entry cohorts. If those cohorts have different event risks, the resulting curve need not represent the target population or a stable follow-up-time process.

Definition: Let \(A\) be calendar date of study entry, \(\tau\) the common administrative study end, \(T\) time from entry to the first event, and \(J\) the event type. Administrative follow-up is \(C=\tau-A\); observed follow-up is \(U=\min(T,C)\). An event is observed when \(T\leq C\), and otherwise follow-up ends administratively. For a competing-risks outcome, \(J\) identifies the cause of an observed event; administrative censoring is not itself an event type.

A shared end does not, by itself, make censoring informative. If entry date is unrelated to the event process and the target population is stable over calendar time, the shorter observation of later entrants may be compatible with standard analyses, subject to adequate follow-up support. But if entry date predicts event risk, or if the event hazard changes with the calendar date at which a person is at risk, then censoring time and the observed event process can be associated in the pooled data.

More precisely, \(C\) is determined by \(A\). Conditional on \(A\), it is fixed, so a formal statement that \(T\) and \(C\) are independent given \(A\) does not by itself settle the analysis question. The event distribution must also be handled appropriately across entry dates. A pooled curve that ignores calendar time can still compare a changing selection of cohorts as follow-up duration increases.

Calendar Time, Follow-Up Time, and the Target

Calendar time and follow-up time are different time scales. A participant entering on calendar date \(A\) and observed \(t\) months later is at calendar date \(A+t\). If the event hazard depends on both time since entry and the calendar date at risk, one way to express the cause-specific hazard is

\lambda_j(t\mid A=a,X=x) = \lim_{\Delta t\to 0} \frac{P(t\leq T<t+\Delta t,\ J=j\mid T\geq t,\ A=a,\ X=x)}{\Delta t},

where \(X\) denotes measured baseline characteristics and \(j\) indexes an event cause. A calendar-time change in treatment, diagnostic practice, eligibility, or background risk can make this hazard depend on \(a\), or on the calendar date \(a+t\), even after accounting for \(t\).

The target must therefore be stated before interpreting a pooled estimate. It might be the event risk by a specified follow-up duration for people entering during a defined calendar period, the risk within each entry cohort, or a standardized risk for a specified distribution of entry dates. These are not automatically the same quantity. In competing risks, the corresponding target may be the cumulative incidence of a particular cause by follow-up duration; other event causes remain events, not administrative censoring.

Key distinction: Administrative censoring describes why observation ends. Calendar-time nonstationarity describes how the event process or the population at risk changes over calendar time. A valid analysis must address both the observation limit and the target event process.

Why the Observed Cohort Mix Changes

At follow-up duration \(t\), a person can be observed only if \(A\leq\tau-t\). Thus the entry dates represented among those with available \(t\)-month outcomes are restricted by the shared end. The following derivation shows why this can matter for a pooled survival curve.

Let \(S(t\mid a)=P(T>t\mid A=a)\), and let \(f_A(a)\) be the entry-date distribution in the target population. The target survival at \(t\), standardized to that population, is

S_{\mathrm{target}}(t)=\int S(t\mid a)f_A(a)\,da.

If we instead average survival only over entry dates with enough administrative follow-up, the mixture is

S_{\mathrm{available}}(t) = \int_{a\leq\tau-t} S(t\mid a) f_A(a\mid A\leq\tau-t)\,da.

These expressions can differ for either of two reasons: survival varies with entry date, or the available entry-date distribution differs from the target distribution. If \(S(t\mid a)\) is constant in \(a\), the change in cohort mix does not change this particular mixture. If survival varies with \(a\), the available-data mixture generally does not equal the target. Estimating a target beyond the follow-up supported by later cohorts requires additional assumptions, not merely a different plotting choice.

Worked Example: How a Shared End Changes the Cohort Mix

Suppose, hypothetically, that 100 early entrants and 100 late entrants are followed until the same study end. Early entrants can contribute 12 months; late entrants can contribute only 6 months. The target is survival at 9 months for the combined entry population, with equal weight for the two cohorts. Assume complete event ascertainment before the study end.

At 9 months, the late cohort has no administrative follow-up support, because \(C=6\) months. Suppose the hypothetical 9-month survival probabilities are 0.90 for early entrants and 0.70 for late entrants. The target survival is

S_{\mathrm{target}}(9) =0.5(0.90)+0.5(0.70) =0.80.

Only the early cohort can contribute observed 9-month outcomes, so its survival of 0.90 is not an estimate of the stated equal-weight target of 0.80. The 0.10 difference arises from both the limited follow-up for late entrants and the assumed cohort difference in survival. If survival were 0.90 in both cohorts, this particular mix difference would not alter the 9-month survival, although the lack of late-cohort follow-up would still limit what can be learned about that cohort at 9 months.

Diagnosing Calendar-Time Trends

Diagnosis begins with a timeline audit, followed by comparisons on scales that separate entry date from follow-up duration. The relevant evidence is not limited to recorded censoring reasons. A study may correctly label every end-of-study observation as administrative while the cohorts entering at different times experience different care or event risks.

Check dates and follow-up support. Confirm entry dates, the common end date, event dates, and the resulting \(C=\tau-A\). Tabulate the number at risk and the maximum supported follow-up by entry cohort.
Compare event experience over a shared follow-up window. For example, compare event risks during the first 6 months after entry for early and late cohorts. This avoids comparing a fully observed early cohort with a late cohort that has not yet reached the same duration.
Inspect calendar-time changes in care and eligibility. Check whether treatment protocols, access to services, diagnostic criteria, enrollment eligibility, or event ascertainment changed during recruitment or follow-up.
Separate causes and populations when relevant. A trend in one event cause may be obscured by a pooled all-event summary. Changes in eligibility can also change who enters, even if care is otherwise stable.
Plot both time scales. Event counts or rates by calendar date can reveal secular changes; event experience by time since entry can show whether cohorts differ at comparable follow-up durations.

A model that includes entry cohort or calendar date can help summarize measured differences, but the form of the model should reflect the question and data. A calendar-time effect may be nonlinear or may change after a policy or care transition. Diagnostics should therefore include plots and substantive review, rather than relying only on a single test of a calendar-time coefficient. Events are observed only while participants remain under observation, so apparent stability among long-follow-up participants cannot establish stability in later cohorts with little follow-up.

Worked Example: Comparing Event Rates on a Common Follow-Up Window

In a hypothetical study with a common end date, 100 early entrants and 100 late entrants each have complete observation for their first 6 months. There are 6 events in the early cohort and 18 in the late cohort during those windows. The total observed person-time is 600 person-months in each cohort. Assume event ascertainment is complete in these windows and the event definitions are unchanged.

The crude event rates are

\widehat{\lambda}_{\mathrm{early}} =\frac{6}{600}=0.0100 \quad\text{events per person-month}, \qquad \widehat{\lambda}_{\mathrm{late}} =\frac{18}{600}=0.0300 \quad\text{events per person-month}.

The late-to-early rate ratio is \(0.0300/0.0100=3.00\). These descriptive rates indicate a substantial cohort difference within the same follow-up window; they do not establish that calendar time caused the difference. Chance variation, changing participant composition, altered care, or altered ascertainment could contribute. The common-window comparison is useful diagnostically because both groups have the same duration of observation, but adjustment or standardization may be needed if their measured characteristics differ.

The rates should not be interpreted as a comparison of 12-month outcomes: the late cohort is not observed that long. Nor should a rate ratio be substituted for a cumulative incidence contrast when competing events or nonconstant hazards are important.

Choosing an Analysis That Matches the Question

The response to a detected trend depends on the target. A calendar-time difference is not automatically a reason to discard the data, but neither does correct administrative censoring justify pooling without checking the event process. Common strategies have different interpretations and limitations.

Approach	What it can address	Main limitation
Restrict to a common follow-up window	Compares cohorts over durations observed for all included groups	Does not answer questions beyond the shortest supported follow-up and may discard information
Report cohort-specific estimates	Shows differences across entry periods directly	Smaller cohorts can yield imprecise estimates; estimates describe different calendar populations
Model or standardize by calendar time and measured covariates	Can address measured differences and define a common target distribution	Depends on correct model specification, measured covariates, and adequate overlap
Pool without calendar adjustment	Provides a simple summary when the event process is sufficiently stable for the stated target	Can mix changing cohort compositions and obscure calendar-time differences

The earlier material on survival and competing risks supplies the usual event-history summaries. Here, the additional requirement is to ask whether those summaries refer to the intended entry population and whether their assumptions remain plausible across calendar time. In particular, if later entrants have no data at a target duration, modeling cannot create empirical support for their unobserved outcomes. An estimate for that duration then relies on extrapolation or assumptions connecting cohorts.

Worked Example: Separating a Common-Window Target From a Longer-Term Target

Consider another hypothetical study with 100 early and 100 late entrants. Both cohorts are completely observed for 6 months. By month 6, 8 early entrants and 20 late entrants have experienced the event. The target is the 6-month risk for a population giving equal weight to these cohorts.

The cohort-specific risks are \(8/100=0.0800\) and \(20/100=0.200\). The equal-weight standardized risk is

\widehat{R}_{\mathrm{target}}(6) =0.5(0.0800)+0.5(0.200) =0.140.

This 6-month estimate is supported for both cohorts, assuming complete event ascertainment and that the specified equal weights define the target. Suppose, however, that only the early cohort can reach 12 months before the study end. A 12-month risk computed from early entrants alone would describe the early cohort, not the equal-weight population. The late-cohort 12-month risk is unobserved, and the equal-weight 12-month risk cannot be calculated from these data without an additional assumption or external information. The analysis should report that limitation rather than silently changing the target population.

Assumptions, Failure Modes, and Common Mistakes

Interpretation of a calendar-time analysis depends on sound data and a clearly specified target. At minimum, verify that dates and event types are recorded correctly; that administrative censoring truly occurs at the stated end; and that outcome ascertainment is not itself changing in a way that creates artificial trends. For adjusted comparisons, relevant differences in entry populations and care must be measured adequately, the model must represent the calendar-time pattern sufficiently well, and there must be enough overlap to compare cohorts at the target times.

Several warning signs indicate that a pooled analysis may be going wrong: event rates differ across entry periods at comparable follow-up; eligibility or care changes coincide with recruitment; the distribution of entry dates among the risk set shifts sharply with follow-up duration; or estimates change materially when the analysis is restricted to a common window or stratified by entry period. Such findings call for investigation and target clarification, not an automatic declaration that the administrative censoring label is incorrect.

Wrong reasoning: “Censoring is administrative, so it is independent and calendar time can be ignored.” Correction: The censoring reason alone does not show that event risks are stable across entry cohorts. Check whether the event process and population change with calendar time.
Wrong reasoning: “The pooled 12-month estimate applies to all entrants because the study followed everyone to the same end date.” Correction: Later entrants may not have 12 months of support. State which cohorts contribute and whether the target includes cohorts without observed outcomes at that duration.
Wrong reasoning: “A calendar-time association proves that the calendar change caused the events.” Correction: It is an observed difference that may reflect changing participant composition, care, ascertainment, chance, or other factors.

Calendar-time adjustment also has limits. An unmeasured change in care or eligibility can remain after adjustment for recorded variables. Sparse data in some periods can make adjusted estimates unstable, while a model may extrapolate across periods with little overlap. Restricting to a common window improves comparability of follow-up duration but does not remove all cohort differences. The defensible choice is the one whose assumptions and target are explicit.

Key takeaway: A shared study end creates a predictable link between entry date and available follow-up. If event risks, care, eligibility, or ascertainment change across calendar time, diagnose those changes and define the target before pooling. Administrative censoring is not a substitute for checking calendar-time comparability.

Check Your Understanding

Answer each question, then compare with the solutions below.

Participants enter at calendar dates \(A\) and are administratively censored at common end \(\tau\). Write their administrative follow-up time.
Why can a pooled 9-month survival estimate fail to represent the full entry population when later entrants have less than 9 months of follow-up?
In a hypothetical common 6-month window, one cohort has 12 events in 600 person-months and another has 24 events in 600 person-months. Calculate both crude rates and their rate ratio.
Name two diagnostics for detecting a calendar-time trend that may affect an event-history analysis.
Does a difference in event rates between calendar periods, by itself, prove that a change in care caused the difference? Explain.

Solutions. (1) \(C=\tau-A\). (2) At 9 months, only entry dates satisfying \(A\leq\tau-9\) can contribute; if survival varies by entry date, the observed cohort mix differs from the target entry population. (3) The rates are \(12/600=0.0200\) and \(24/600=0.0400\) events per person-month; the second-to-first rate ratio is \(0.0400/0.0200=2.00\). (4) Examples include comparing event experience over a common follow-up window by entry cohort, plotting rates by calendar date, tabulating follow-up support by entry period, and reviewing changes in care or eligibility. (5) No. The difference may reflect participant composition, care, ascertainment, chance, or other factors; a descriptive calendar-time association does not establish causation.