Censoring in Survival Analysis

The Problem of Incomplete Data: Censoring

In an ideal study, we would follow every subject from a starting point until the event of interest (e.g., death, policy lapse, equipment failure) occurs. We would then have a complete set of event times. In reality, this is rarely possible. For many subjects, we only have partial information.

Censoring occurs when we have some information about an individual's survival time, but we do not know the exact time of the event. For example, a study might end before all participants have died, or a participant might move away and be lost to follow-up. We can't simply discard these individuals, as that would discard valuable information and likely bias our results.

To formalize this, let's establish some key notation which will be essential throughout this module. For each individual $i$:

• Let $T_i$ be the true, underlying time to the event of interest.
• Let $C_i$ be the censoring time (e.g., the time they drop out of the study or the time the study ends).

In a typical study, we don't observe both $T_i$ and $C_i$. We only observe the time of the first thing that happens: either the event or censoring. We define the observed time as:

To know what happened at time $Y_i$, we also need an event indicator, often denoted by ${\delta }_i$:

Our data for survival analysis will therefore consist of pairs of $(Y_i,{\delta }_i)$ for each individual.

The following video provides an excellent introduction to this topic, framing it within the broader context of statistical learning and illustrating why censoring is a critical feature of time-to-event data.

Types of Censoring

Censoring can occur in different ways, depending on what information is missing. The three primary types are right, left, and interval censoring. The following video gives a clear and practical explanation of each.

Let's explore these types in more detail, supplementing the video with more formal definitions and examples.

1. Right Censoring

This is by far the most common type of censoring in survival analysis. Right censoring occurs when the event of interest has not yet happened by the end of the observation period. We only know that the true event time $T$ is greater than some observed time $Y$.

Common causes include:

• End of study: The study concludes at a pre-determined date, and some subjects are still event-free.
• Loss to follow-up: A subject withdraws from the study, moves away, or can no longer be contacted for other reasons.
• Competing event: The subject experiences a different event that removes them from being at risk for the event of interest (e.g., in a study on death from cancer, a patient dies in a car accident).

The zedstatistics video provides a great walkthrough of right censoring in a real-world cancer study.

2. Left Censoring

Left censoring is the mirror image of right censoring. It occurs when the event of interest has already happened before the observation period begins. We only know that the true event time $T$ is less than some observation time.

This is less common, particularly in mortality studies, because you typically can't enroll a subject into a study if they are already dead. However, it can occur in other contexts. For example, in a study of the age at which a virus is contracted, some participants may already test positive at their first screening. We know they were infected before that screening, but we don't know exactly when.

3. Interval Censoring

Interval censoring occurs when we know that the event happened within a specific time interval, but we don't know the exact moment. This is common in studies where subjects are observed only at periodic follow-up appointments.

For example, if a patient is tested for a disease annually, and they test negative in 2022 but positive in 2023, we know the disease onset occurred sometime between those two appointments. The true event time $T$ is in the interval $(t_{2022},t_{2023}]$.

The image below provides a helpful visual summary of these censoring types for several individuals in a hypothetical study.

An Important Distinction: Censoring vs. Truncation

It's important to distinguish censoring from a related concept called truncation.

• With censored data, we have an incomplete observation for an individual who is included in our sample.
• With truncated data, certain individuals are systematically excluded from the sample altogether due to the study design. We have no information about them at all.

Left truncation is the most common form. This happens when subjects are only included in a study after they have experienced some intermediate event. For example, a study on survival after a heart transplant only includes patients who survive long enough to receive a transplant. Patients who die while on the waiting list are left-truncated—they are never part of the study sample.

For a deeper dive into these definitions, including the subtypes of right censoring and the distinction with truncation, the following resource is very useful.

The Assumption of Non-Informative Censoring

The statistical methods we will learn to handle censored data, like the Kaplan-Meier estimator in our next lesson, rely on a crucial assumption: non-informative censoring. This assumption states that the reasons for an individual being censored are not related to their probability of experiencing the event of interest.

For example, if patients who are sicker and more likely to die soon are also more likely to drop out of a study (perhaps because they are too unwell to attend appointments), the censoring is informative. Simply treating these dropouts as right-censored would lead to an overestimation of survival times, as we would be systematically removing the highest-risk individuals from the analysis.

In most cases, we must assume that censoring is non-informative. For instance, administrative censoring (where the study ends) is generally non-informative. However, determining whether loss to follow-up is informative often requires careful investigation of the study's design and is not something that can be tested purely statistically. This concept is mentioned in the CS2 exam materials, making it an important point to understand.

The following segment from the Stanford video provides a clear discussion of this potential bias.

Conclusion

In this lesson, we have moved from the world of theoretical probability models to the practical realities of working with survival data. You've learned how to identify the different ways data can be incomplete and why these situations require special handling.

Key Takeaways:

• Censoring occurs when the time-to-event is not known exactly for a subject. The data for a censored observation consists of an interval (lower or upper bound) and an indicator variable.
• The main types of censoring are right censoring (event occurs after observation ends), left censoring (event occurs before observation begins), and interval censoring (event occurs between two observation points).
• Truncation is different from censoring; it means that some individuals are entirely excluded from the sample due to the study design.
• Most survival analysis methods rely on the assumption of non-informative censoring, meaning that the act of censoring is independent of the subject's prognosis.

With this foundational understanding of censored data, you are now prepared to learn how to analyze it. In our next lesson, we will construct the Kaplan-Meier estimator, a powerful non-parametric method for estimating the survival function directly from data containing right-censored observations.