Censoring and Follow-Up Data

Censoring is the defining feature of time-to-event data: for some subjects the event of interest has not occurred by the end of observation, so their true event time is unknown and known only to lie beyond their last recorded follow-up. Handling this partial information correctly — rather than discarding incompletely observed subjects — is what makes survival analysis distinct from ordinary statistical methods.

Definition

Censoring is incomplete observation of an event time: a subject is right-censored when they are event-free at their last follow-up, left-censored when the event is known to have occurred before observation began, and interval-censored when it is known only to fall between two assessment times.

Scope

This topic covers the types of censoring (right, left, interval) and truncation, the assumptions that make censored data usable — chiefly that censoring is non-informative — and the role of follow-up time and loss to follow-up. It is a methodological reference and does not address clinical management of individual patients.

Core questions

• What does it mean for an observation to be censored, and what are the principal types of censoring?
• Why can censored subjects not simply be deleted from the analysis?
• What is the assumption of non-informative (independent) censoring, and when might it fail?
• How do follow-up time and loss to follow-up affect the validity of survival estimates?

Key concepts

• Right censoring
• Left censoring
• Interval censoring
• Left truncation (delayed entry)
• Administrative censoring
• Non-informative (independent) censoring
• Loss to follow-up
• At-risk time and person-time

Mechanisms

When the event is not observed for a subject, their record still contributes information up to the moment they were last known to be event-free; survival methods use this by counting the subject in the risk set until their censoring time. The key requirement is non-informative censoring: the reason a subject is censored must be unrelated to their underlying risk of the event, so that those remaining under observation represent those who were censored. When this fails — for example, when sicker patients are preferentially lost to follow-up — survival estimates become biased. Right censoring (event-free at study end or at dropout) is by far the most common form in medical research; left and interval censoring arise when the event timing is only partially known (Leung et al., 1997; Clark et al., 2003).

Clinical relevance

Whether a study's follow-up was complete and whether dropout was related to prognosis are central to appraising any survival result, since informative censoring can distort reported survival and treatment effects. This topic explains why those features matter; it describes analytic considerations and is not clinical guidance.

Epidemiology

Censoring is ubiquitous in cohort studies and clinical trials with finite follow-up; loss to follow-up is a recognised threat to validity, and incomplete follow-up is routinely reported and scrutinised in epidemiologic and trial publications (Leung et al., 1997).

Evidence & guidelines

There are no clinical guidelines for censoring; the reference standards are foundational statistical work and biostatistics texts. Kaplan and Meier (1958) framed estimation from incomplete observations, and texts such as Klein and Moeschberger (2003) and Collett (2015) treat censoring and truncation systematically, while reporting standards for trials and cohorts emphasise complete and unbiased follow-up.

History

The treatment of incompletely observed lifetimes has roots in actuarial life tables, but the formal statistical handling of censored data was consolidated alongside the Kaplan-Meier estimator in 1958, whose title — 'Nonparametric Estimation from Incomplete Observations' — names the problem directly. Subsequent texts elaborated the taxonomy of censoring and truncation and the independence assumptions on which valid inference depends (Klein & Moeschberger, 2003).

Debates

When is censoring informative, and how should it be addressed?

Standard methods assume censoring is independent of event risk; when dropout is related to prognosis this assumption fails, biasing estimates, and there is ongoing methodological work on sensitivity analyses and models for dependent censoring.

Key figures

• Edward L. Kaplan
• Paul Meier
• John P. Klein
• Melvin L. Moeschberger

Seminal works

• kaplan-meier-1958
• leung-1997

Frequently asked questions

Why not just exclude subjects whose event was never observed?

Excluding censored subjects discards real information about how long they remained event-free and biases the result toward those who had the event; survival methods instead keep censored subjects in the risk set up to their last follow-up.

What is non-informative censoring?

It is the assumption that a subject's reason for being censored is unrelated to their underlying risk of the event, so that those still under observation fairly represent those who were censored; if it fails, survival estimates can be biased.