Administrative Censoring

Administrative Censoring, Immature Survival Data and Extrapolation

Every trial that feeds an economic model stops collecting outcome data on some date, leaving anyone still event-free with an incomplete survival time. In health technology assessment, data in which many patients are still event-free at the cut-off are described as immature, and the proportion still event-free at the end of follow-up, together with the hazard assumed beyond it, largely drives how much of mean survival, and so of the QALY gain, must come from extrapolation rather than observation. Censoring times arise from staggered recruitment and are normally treated as non-informative, an assumption that can fail, and they shape the Kaplan-Meier curve and the restricted mean before carrying into partitioned survival and state transition models. An illustrative eight-patient example works through each calculation.

How a data cut-off creates censoring times

Patients enter a trial over months or years, but their outcomes are analysed at a single calendar date, the data cut-off, set for a planned interim or final analysis. Each patient's potential follow-up is the gap between their own entry date and that shared date. Early recruits are therefore observed for longer than late ones, as the definitions below show.

$$C_i = d - e_i, \qquad Y_i = \min(T_i, C_i), \qquad \delta_i = I(T_i \le C_i)$$

where $C_i$ is the administrative censoring time of patient $i$ measured from entry or randomisation, $d$ is the calendar date of the cut-off, $e_i$ is the calendar date of entry, $Y_i$ is the observed follow-up time, $T_i$ is the true (possibly unobserved) time to the event, and $\delta_i$ equals 1 if the event was observed by the cut-off and 0 if the patient was censored. When patients can also be lost to follow-up, the censoring time becomes $C_i = \min(d - e_i, L_i)$, where $L_i$ is the time at which patient $i$ left observation for another reason; only the first component is administrative.

The potential follow-up $d - e_i$ is fixed by the study timetable for every patient, including those who have the event, so it is a property of the design rather than of patient behaviour, and randomisation spreads entry dates across arms.

Why administrative censoring is usually treated as non-informative

Standard survival methods are valid only when censoring is non-informative, meaning that a patient censored at a given time should be as likely to have a later event as one who remains under observation. Administrative censoring usually meets this condition because the censoring time depends on the date of entry, not on prognosis. Formally $T_i$ must be independent of $C_i$, and because $C_i$ is set by $e_i$, the condition reduces to the requirement that time to event not depend on when the patient was recruited.

When staggered entry breaks the assumption

That reduction shows where the assumption fails. Clark and colleagues note that the Kaplan-Meier method assumes survival probabilities are the same for patients recruited early and late, and that changes in treatment, earlier diagnosis or other changes over time can produce spurious results. If later lines of therapy improve during a long recruitment period, late recruits, who supply only short follow-up, differ in prognosis from early recruits, who alone supply the tail, and any extrapolation fitted to that tail reflects the earliest cohort. Comparing survival by recruitment period over the follow-up time the periods share is a simple check.

Induced informative censoring for costs and QALYs

Censoring that is harmless for survival time can still bias cumulative cost or quality-adjusted survival. Two patients censored after the same follow-up have usually accrued different costs and quality-adjusted time, and the amount depends on their health. Lin and colleagues showed that the naive average of observed costs, over all patients or only uncensored ones, can be severely biased and that standard survival techniques are not applicable to costs; they proposed estimators that combine Kaplan-Meier probabilities with costs in short intervals. The NICE Decision Support Unit review of partitioned survival analysis (TSD 19) traces the same problem for quality-adjusted survival to Goldhirsch, Glasziou and colleagues, who noted in 1989 and 1990 that standard survival methods induced informative censoring because people with worse health-related quality of life were more likely to be censored earlier.

Effect on the Kaplan-Meier curve and the restricted mean

The Kaplan-Meier estimator uses each censored patient for as long as they are observed and then removes them from the risk set. Administrative censoring removes patients in a predictable order, latest recruits first. The number at risk therefore falls steadily towards the end of follow-up, as the estimator below reflects.

$$\hat{S}(t) = \prod_{j:\, t_j \le t} \left(1 - \frac{d_j}{n_j}\right), \qquad \text{RMST}(\tau) = \int_0^{\tau} \hat{S}(t)\, dt$$

where $\hat{S}(t)$ is the estimated probability of remaining event-free beyond time $t$, $t_j$ are the distinct event times, $d_j$ is the number of events at $t_j$, $n_j$ is the number still at risk just before $t_j$ (excluding anyone already censored), and $\text{RMST}(\tau)$ is the restricted mean survival time up to a restriction time $\tau$ no later than the last observed follow-up.

Because only early recruits remain at risk late in follow-up, the final steps rest on few patients, and Clark and colleagues note that confidence limits are often wide at the tail. The NICE Decision Support Unit's TSD 14 adds that when some patients remain alive at the end of follow-up, only a mean restricted to that time point can be estimated directly. Maturity is usually summarised by the proportion with the event and by median follow-up, estimated with the reverse Kaplan-Meier method, in which the event indicator is reversed so that being censored becomes the outcome.

From censored trial data to partitioned survival and state transition models

Economic models usually need mean survival over a lifetime time horizon. TSD 14 states that unless trial data are complete or very close to complete, extrapolation is required, and that parametric models are likely to be the preferred method in most cases. TSD 21 observes that oncology trials are usually subject to a large amount of administrative censoring, with short follow-up relative to patients' lifespans.

In a partitioned survival model, state membership comes directly from curves such as overall survival and progression-free survival. TSD 19 explains that the overall survival curve may be extrapolated beyond the study depending on the time horizon and the maturity of the data, that the area under the extrapolated curve gives mean life expectancy, and that without external information each endpoint's extrapolation reflects only its own within-trial trend. In a state transition model, extrapolated survival depends instead on the model structure and the estimated transition probabilities, so administrative censoring acts through the data behind each transition.

The NICE manual (PMG36, sections 4.2.24 and 4.6.25) says a lifetime horizon often requires extrapolating beyond the available studies, asks for the clinical plausibility of the underlying hazards to be assessed, and warns that functions with stable or decreasing variance over time are likely to understate uncertainty in the extrapolation. TSD 21 adds that AIC and BIC say little about how well a model extrapolates. TSD 14 suggests that where censoring is substantial a scenario based only on observed data can show how far the extrapolated period drives the incremental cost-effectiveness ratio.

Data cut-offs in HTA submissions

A health technology assessment submission rests on whichever data cut was available. A later cut moves every administrative censoring time forward, turns some censored observations into events and lengthens the observed tail, so an earlier extrapolation can be tested against new data. In one simulated dataset in TSD 21, drawn from a scenario in which many methods performed poorly, extending administrative censoring from 3 to 5 years improved the extrapolation from a model incorporating background mortality but worsened the mean survival estimate from an all-cause model, and the authors conclude that hazard projections must still be credible with longer follow-up.

Where follow-up is short, the NICE manual (PMG36, section 5.5.19) lets a company include a managed access proposal for medicines where immature evidence or evidence gaps are likely to cause significant uncertainty for committee decision making. Under sections 6.4.6 and 6.4.9, a recommendation with managed access, a form of managed entry agreement, makes the medicine available for a time-limited period while further evidence, for example from ongoing trials, is collected, so a later data cut can address the uncertainty that administrative censoring created.

Worked example with eight patients and a 36-month cut-off

The figures are illustrative. One trial arm recruits eight patients between month 0 and month 24, and the data cut-off falls at month 36 of the trial calendar. No patient is lost to follow-up, so every censored observation is administrative.

1. Censoring times. Each censored patient's observed time equals the cut-off minus their entry month; for the last recruit, $C_8 = 36 - 24 = 12$ months.

2. Kaplan-Meier estimate. Eight patients are at risk at month 6 and seven at month 10, with one death each time. Patient 8 is censored at month 12 and patient 7 at month 16, leaving five at risk at month 14 and three at month 20 in the estimate that follows.

$$\hat{S}(6) = 1 - \frac{1}{8} = 0.875, \qquad \hat{S}(10) = 0.875 \times \left(1 - \frac{1}{7}\right) = 0.75, \qquad \hat{S}(14) = 0.75 \times \left(1 - \frac{1}{5}\right) = 0.6, \qquad \hat{S}(20) = 0.6 \times \left(1 - \frac{1}{3}\right) = 0.4$$

where $\hat{S}(t)$ is the estimated survival probability after month $t$ and each factor is one minus the deaths divided by the number at risk. Patients 5 and 2 are censored at months 24 and 34 without further deaths, so the curve stays at 0.40 to month 34 and is undefined beyond it. After month 24 it rests on one patient, the second recruit.

3. Naive summaries compared. Half the patients were alive at the cut-off, but that ignores the shorter follow-up of late recruits; dropping the censored patients leaves four known outcomes, all deaths, implying no survivors. The Kaplan-Meier estimate of 0.40 uses each patient only for as long as they were observed.

4. Restricted mean to month 34. The area under the curve sums survival multiplied by interval width over 0 to 6, 6 to 10, 10 to 14, 14 to 20 and 20 to 34 months, as follows.

$$\text{RMST}(34) = 1 \times 6 + 0.875 \times 4 + 0.75 \times 4 + 0.6 \times 6 + 0.4 \times 14 = 21.7$$

where $\text{RMST}(34)$ is the restricted mean survival time in months.

5. The tail beyond the last observed follow-up. With 40% still alive at month 34, the lifetime mean depends on what is assumed afterwards. A constant hazard estimated from the trial, and the area added by a constant hazard after month 34, are given below.

$$\hat{h} = \frac{\sum_i \delta_i}{\sum_i Y_i} = \frac{4}{136} = 0.0294, \qquad \int_{34}^{\infty} 0.4\, e^{-h(t - 34)}\, dt = \frac{0.4}{h}$$

where $\hat{h}$ is the estimated monthly hazard (deaths divided by total patient-months observed), $h$ is the hazard assumed after month 34 and 0.4 is survival at month 34. Three illustrative tail assumptions give the results in the table.

6. Interpretation. Between about a third and a half of the mean comes from after month 34, the last observed follow-up, and eight patients cannot discriminate between these assumptions, so the arm's mean survival, and hence the incremental QALY gain once a comparator is added, can move substantially with the tail assumption. A data cut at month 48 of the trial calendar would add up to 12 months of observation for each of the four patients still alive, extending the curve to as far as month 46. If prognosis had improved during recruitment, the curve beyond month 24, resting on patient 2, one of the earliest recruits, would misstate survival for the arm. A real analysis would compare several parametric forms and external evidence.

Administrative censoring compared with other censoring

Administrative censoring is one source of right censoring in survival analysis, and each source carries a different risk of bias. The table sets out the main contrasts.

The censoring indicator alone cannot show which mechanism applied, so an analysis relying on non-informative censoring needs to know how many censored patients reached the cut-off and how many left early.

Where administrative censoring is mishandled

The commonest error is to summarise immature data with simple proportions, counting censored patients as survivors or dropping them, as the worked example shows. Both ignore the unequal follow-up that staggered recruitment produces.

A second error is to read the flat final segment of a Kaplan-Meier curve as a plateau or cure. After the last event the curve is flat by construction, and TSD 21 cautions that when few patients remain at risk towards the end of follow-up, as is common in randomised trials, assuming cure is highly questionable; a cure model fitted on that basis can extrapolate very poorly. Further errors are reporting median follow-up as the median of observed times, which counts early deaths as short follow-up; assuming that censoring which is non-informative for survival is also non-informative for costs and QALYs; and confusing the data cut-off, which limits what was observed, with the model's time horizon, which is an analytic choice.

Sources

• Clark TG, Bradburn MJ, Love SB, Altman DG. Survival analysis part I: basic concepts and first analyses. British Journal of Cancer. 2003;89(2):232-238.
• Kalbfleisch JD, Prentice RL. The Statistical Analysis of Failure Time Data. 2nd ed. Hoboken, NJ: Wiley; 2002.
• Latimer N. NICE DSU Technical Support Document 14: Survival analysis for economic evaluations alongside clinical trials, extrapolation with patient-level data. Sheffield: Decision Support Unit, ScHARR; 2011, last updated March 2013.
• Lin DY, Feuer EJ, Etzioni R, Wax Y. Estimating medical costs from incomplete follow-up data. Biometrics. 1997;53(2):419-434.
• National Institute for Health and Care Excellence. NICE technology appraisal and highly specialised technologies guidance: the manual (PMG36). London: NICE; 2022, updated March 2026. Sections 4.2.24, 4.6.25, 4.6.26, 5.5.19, 6.4.6 and 6.4.9.
• Rutherford MJ, Lambert PC, Sweeting MJ, Pennington B, Crowther MJ, Abrams KR, Latimer NR. NICE DSU Technical Support Document 21: Flexible methods for survival analysis. Sheffield: Decision Support Unit, ScHARR; 2020.
• Woods B, Sideris E, Palmer S, Latimer N, Soares M. NICE DSU Technical Support Document 19: Partitioned survival analysis for decision modelling in health care: a critical review. Sheffield: Decision Support Unit, ScHARR; 2017.