Survival Analysis

The defining feature of survival analysis is its ability to incorporate censored observations—patients who haven't experienced the event by study end or are lost to follow-up. Right censoring is most common: we know the survival time exceeds a certain value but not the exact time. Left censoring occurs when the event happened before observation began. Interval censoring means the event occurred between two observation points. Ignoring censored data by treating it as event-free or excluding it introduces systematic bias, underestimating true survival times.

The Kaplan-Meier (KM) estimator is the non-parametric gold standard for estimating the survival function from lifetime data. It calculates the probability of surviving past a given time point by multiplying conditional survival probabilities at each observed event time. Key properties:

• Produces the characteristic step function plot that drops only at event times
• Handles right-censored data without parametric assumptions
• Allows visual comparison of survival curves between groups using the log-rank test
• Assumes censoring is non-informative—censored patients have the same future risk as those who remain

The Cox model is the workhorse of survival analysis, relating covariates to the hazard rate without specifying the baseline hazard function. It assumes proportional hazards—the effect of a predictor is constant over time. The model expresses the hazard as:

h(t|X) = h₀(t) × exp(β₁X₁ + β₂X₂ + ... + βₚXₚ)

Where h₀(t) is the unspecified baseline hazard and exp(βᵢ) represents the hazard ratio for a one-unit increase in covariate Xᵢ. A hazard ratio of 2.0 means the event rate doubles. Violations of the proportional hazards assumption require time-varying coefficients or stratified models.

In clinical settings, patients may experience mutually exclusive events that prevent the primary event of interest. For example, death from cardiovascular causes precludes observing death from cancer. Standard survival methods that censor competing events produce biased estimates by treating them as independent. The Fine-Gray subdistribution hazard model and cause-specific hazard models explicitly account for competing risks. The cumulative incidence function (CIF) estimates the probability of experiencing a specific event type by a given time, accounting for all competing events simultaneously.

Many clinical predictors change during follow-up—lab values fluctuate, treatments switch, and disease states progress. Time-dependent covariates capture these dynamic exposures. The extended Cox model incorporates variables that update at each event time:

• Internal covariates: Values generated by the patient's own disease process (e.g., CD4 count in HIV)
• External covariates: Values determined independently of the patient's health trajectory (e.g., air pollution levels)

Proper handling requires data restructuring into counting process format with multiple rows per patient, each representing a time interval with constant covariate values.

Clinical events like hospital readmissions, infections, or tumor recurrences can occur multiple times per patient. Standard survival methods assuming independent observations produce overly narrow confidence intervals by ignoring within-subject correlation. Specialized models include:

• Andersen-Gill model: An intensity-based counting process treating events as conditionally independent given covariates
• Prentice-Williams-Peterson models: Stratified approaches that reset the time clock after each event
• Frailty models: Introduce random effects to account for unobserved heterogeneity between patients

These methods are critical for evaluating interventions in chronic disease management where repeated events are the norm.

A non-parametric statistic used to estimate the survival function from lifetime data. It calculates the probability of surviving past a certain time point, even when patients are censored (lost to follow-up or haven't experienced the event by the study's end). The resulting Kaplan-Meier curve visually depicts the proportion of a population still event-free over time, with characteristic step-wise drops at each event occurrence.

A regression model that investigates the effect of multiple variables on the hazard rate—the instantaneous risk of an event occurring at time t, given survival up to t. It assumes the hazard ratios for different predictors are constant over time. The model outputs a hazard ratio for each covariate, quantifying its multiplicative effect on the baseline hazard. For example, a hazard ratio of 2.0 for a treatment group means the event rate is twice that of the control group at any given moment.

A defining feature of survival data where the exact event time is unknown. Right censoring occurs when a subject leaves the study before an event or the study ends without the event occurring. Left censoring happens when the event occurred before study entry. Interval censoring means the event is known only to have occurred within a time window. Properly handling censoring is critical to avoid attrition bias and ensure unbiased survival estimates.

Predictor variables whose values change over the observation period, violating the standard Cox model's proportional hazards assumption. Examples include repeated lab measurements, medication adherence, or a patient crossing over to a new treatment. Modeling these requires extended Cox models or landmark analysis, which resets the baseline at specific time points to assess the effect of updated covariate values on subsequent survival.

An extension of survival analysis for scenarios where a subject can experience one of several mutually exclusive events, and the occurrence of one precludes the others. For example, in an oncology study, death from cancer progression and death from treatment toxicity are competing risks. The Cumulative Incidence Function (CIF) estimates the probability of a specific event over time, accounting for the competing events, unlike the 1-KM estimator which would overestimate risk.

A non-parametric hypothesis test used to compare the survival distributions of two or more groups. It tests the null hypothesis that there is no difference between the populations in the probability of an event at any time point. The test compares observed versus expected event counts across all time points. It is most powerful when hazard ratios are proportional over time and is the standard statistical test accompanying Kaplan-Meier curve comparisons.