Censoring and Survival Functions
Time-to-event targets, incomplete observation, and the risk process
1 Event Time and Observation Time Are Different Variables
Survival analysis studies the time \(T\geq 0\) from a defined origin to a defined event. The event may be death, relapse, device failure, cell differentiation, or another first occurrence. Its name does not have to imply mortality. What makes the analysis distinctive is that follow-up may end before the event is observed.
Under ordinary right censoring, let \(C\) be the censoring time and record
\[ \boxed{X=\min(T,C),\qquad \Delta=I(T\leq C).} \]
\(X\) is the observed follow-up time and \(\Delta\) says whether that time is an event. A censored record with \((X,\Delta)=(12,0)\) means \(T>12\); it does not mean that the event occurred at 12, nor that the event time is missing without information.
The time origin, event definition, and observation process are therefore part of the estimand. Overall survival from diagnosis, progression-free survival from randomization, and time to recurrence after surgery are different variables even when measured on the same people.
2 Survival, Distribution, Density, and Hazard
For a continuous event time with distribution function \(F\) and density \(f\), the survival function is
\[ S(t)=P(T>t)=1-F(t). \]
\(S(t)\) is a population probability: the probability that a randomly selected unit from the target population remains event-free beyond \(t\). It is nonincreasing, right-continuous, and lies between zero and one. The \(p\)th survival quantile is the time at which survival falls to \(1-p\); the median is not estimable within the observed range if the estimated curve never reaches \(0.5\).
The hazard describes instantaneous event intensity among units still at risk:
\[ h(t)=\lim_{\Delta t\downarrow 0} \frac{P(t\leq T<t+\Delta t\mid T\geq t)}{\Delta t} =\frac{f(t)}{S(t)}. \]
It is a rate, not a probability. It may exceed one and depends on the time unit. The cumulative hazard and survival function satisfy
\[ H(t)=\int_0^t h(u)\,du, \qquad \boxed{S(t)=\exp\{-H(t)\}}. \]
This identity follows from \(h(t)=-d\log S(t)/dt\). It connects models for instantaneous risk with probabilities over a time interval, but it does not make a hazard ratio a survival-probability ratio.
When the mean exists, restricted follow-up has a particularly interpretable summary:
\[ \operatorname{RMST}(\tau)=E\{\min(T,\tau)\} =\int_0^\tau S(t)\,dt. \]
This is the expected event-free time accumulated through a prespecified horizon \(\tau\). Unlike an unrestricted mean, it can be estimated without extrapolating beyond that horizon when follow-up supports it.
3 How Observation Can Be Incomplete
| Mechanism | What is learned about \(T\) | Example |
|---|---|---|
| Right censoring | \(T>C\) | study closes before relapse |
| Left censoring | \(T\leq C\) | infection occurred before first positive visit |
| Interval censoring | \(L<T\leq R\) | conversion occurred between scheduled assays |
| Left truncation (delayed entry) | unit is observed only if \(T>E\) | registry enrollment after diagnosis |
| Competing event | another event prevents the event of interest | death before relapse |
Censoring limits what is observed after inclusion. Truncation determines who enters the observed sample. A person entering at time \(E\) must join the risk set at \(E\), not at the original time zero. Treating delayed entry as ordinary right censoring creates immortal-time bias because the person had to remain event-free long enough to be observed.
Administrative censoring at a common study-closing date is often plausibly unrelated to prognosis, although staggered enrollment makes censoring times vary. Loss to follow-up is more concerning: illness, toxicity, treatment response, socioeconomic conditions, or sequencing failure may influence both dropout and the event time.
A competing event is not simply benign censoring when the target is the real-world probability of the event of interest. For example, censoring death while estimating relapse incidence imagines a world in which those who died could remain at risk for relapse. Cause-specific hazards and cumulative incidence answer different questions and require a competing-risks framework.
4 Likelihood Contributions Explain the Information
Suppose right censoring is independent of the event process under the model. Ignoring factors that contain only parameters of the censoring distribution, unit \(i\) contributes
\[ L_i(\theta) =f_\theta(x_i)^{\delta_i}S_\theta(x_i)^{1-\delta_i} =\{h_\theta(x_i)\}^{\delta_i}S_\theta(x_i). \]
An event contributes density, or equivalently hazard times survival to the event time. A censored observation contributes the probability of surviving beyond its last observed time. Replacing censored times by endpoints or deleting them changes this likelihood and usually biases the observed distribution toward short times.
The common phrase non-informative censoring is shorthand for a conditional assumption, not a property that a censoring mark reveals. A useful version is
\[ T\perp C\mid Z, \]
where \(Z\) contains the design variables and prognostic information used to make the assumption credible. Marginal independence may fail even when conditional independence is plausible. Conversely, conditioning on variables affected by treatment or impending failure can create a different problem.
The assumption cannot be proved from the observed \((X,\Delta)\) alone because the post-censoring event times are precisely what are unavailable. Design knowledge, reasons for censoring, follow-up procedures, auxiliary data, and sensitivity analysis are essential.
5 Estimands, Estimators, and Estimates
Survival data support several distinct targets:
| Scientific question | Possible estimand | Typical method |
|---|---|---|
| What fraction remains event-free at \(t\)? | \(S(t)\) | Kaplan–Meier under independent right censoring |
| How much event-free time accumulates through \(\tau\)? | \(\int_0^\tau S(t)dt\) | restricted mean survival time |
| What is the instantaneous event intensity? | \(h(t)\) or a hazard contrast | hazard estimator or regression model |
| What is the real-world probability of a particular event by \(t\)? | cumulative incidence | competing-risks estimator |
| How does a covariate alter the event-time distribution? | prespecified survival, hazard, or time contrast | Cox or parametric survival model |
The estimand is the population quantity. A product-limit curve, regression procedure, or parametric fit is an estimator. Its realized curve or coefficient is an estimate. A p-value from a group comparison is neither an effect estimate nor a description of survival at a clinically relevant time.
6 Design and Bioinformatics Interpretation
Consider a cancer-genomics cohort evaluating time from surgery to recurrence. The patient is the biological and inferential unit. Multiple tumor regions, aliquots, libraries, reads, or cells improve measurement of a patient’s molecular state but do not create additional independent survival times.
Before modeling, specify:
- the target population and eligibility criteria;
- the time origin and whether entry is delayed;
- the event definition and adjudication rule;
- the analysis horizon and competing events;
- the censoring reasons and last-contact rule;
- whether a biomarker was prespecified or selected using the same outcomes;
- biological units, technical replicates, batches, and missing assays.
Statistically, an estimated survival difference describes the chosen endpoint under the sampling and censoring assumptions. Biologically, it does not by itself show that a molecular signature causes recurrence. Stage, treatment, ascertainment, cohort selection, batch effects, and data-dependent signature construction may explain an association. Technical replication can reduce measurement error; only independent patients support population-level replication.
7 Diagnostics and Sensitivity
| Question | Evidence to inspect | Sensitivity or remedy |
|---|---|---|
| Are time zero and eligibility aligned? | enrollment and origin dates; impossible negative or zero intervals | rebuild start–stop records; use delayed entry |
| Could censoring depend on prognosis? | reasons and timing by treatment, biomarker, site, and health status | adjust for predictors; inverse-censoring weights; pattern or tipping-point analyses |
| Is the tail supported? | numbers at risk and censoring marks | restrict the horizon; avoid unsupported medians or extrapolation |
| Are event definitions consistent? | adjudication and coding by site or batch | harmonize endpoints; analyze alternative defensible definitions |
| Are there competing events? | event-type counts over time | report cumulative incidence alongside cause-specific analyses |
| Is dependence represented? | repeated events, paired organs, families, sites | cluster-aware variance or multilevel/event-process model |
Comparing the measured baseline characteristics of censored and uncensored units can reveal warning signs but cannot validate independent censoring. A useful sensitivity analysis states how departures from the assumption would alter the conclusion rather than declaring censoring harmless because one diagnostic test was nonsignificant.
8 Common Pitfalls
- Calling censoring missing data with no information: right censoring gives a lower bound on \(T\) and must enter the risk process.
- Treating a hazard as a probability: a hazard is an instantaneous rate conditional on remaining at risk.
- Ignoring delayed entry: including a late entrant in earlier risk sets credits the group with survival time during which that unit could not fail in the observed sample.
- Reporting a median that was not reached: the curve must cross \(0.5\) within adequately observed follow-up.
- Censoring competing events without naming the hypothetical target: this does not estimate the ordinary cumulative probability of the event.
- Counting technical measurements as survival replicates: the independent biological unit, not the number of reads or cells, determines replication.
9 Connections and Take-home Messages
- Random variables and probability distributions defines distribution and survival functions.
- Exponential and geometric distributions develops constant hazards and memorylessness.
- Likelihood and log-likelihood explains why an observation mechanism changes likelihood contributions.
- Repeated measures and longitudinal data separates correlated trajectories from time-to-first-event outcomes.
- Sensitivity analysis and robustness provides a framework for claims under unverifiable assumptions.
- Survival analysis models an event time and the process that makes it observable.
- \(S(t)\), \(h(t)\), and \(H(t)\) are linked but answer different questions.
- Independent censoring is conditional and assumption-dependent; it is not certified by a censoring indicator.
- Censoring, truncation, and competing events require different risk sets and target definitions.
- Bioinformatics survival claims are replicated by independent biological units, not by technical subsamples.
10 References
- Kalbfleisch, J. D., and Prentice, R. L. (2002). The Statistical Analysis of Failure Time Data (2nd ed.). Wiley. https://www.wiley.com/en-ca/the-statistical-analysis-of-failure-time-data-2nd-edition-p-9781118032985
- Andersen, P. K., Borgan, O., Gill, R. D., and Keiding, N. (1993). Statistical Models Based on Counting Processes. Springer. https://doi.org/10.1007/978-1-4612-4348-9
- Kaplan, E. L., and Meier, P. (1958). Nonparametric estimation from incomplete observations. Journal of the American Statistical Association, 53(282), 457–481. https://doi.org/10.1080/01621459.1958.10501452
- Aalen, O. O., and Johansen, S. (1978). An empirical transition matrix for non-homogeneous Markov chains based on censored observations. Scandinavian Journal of Statistics, 5(3), 141–150. https://www.jstor.org/stable/4615704