In classical survival analysis, we track individuals over time to see when a specific event of interest occurs—such as the recurrence of a disease or the failure of a mechanical part. If a participant leaves a study early or the study ends before their event occurs, we classify them as "censored."
Standard survival models, like the Kaplan-Meier estimator or the Cox proportional hazards model, make a fundamental assumption about censoring: it must be independent or informative-neutral. This means that a censored individual is assumed to face the exact same risk of experiencing the event in the future as those who remain under observation.
However, in real-world clinical settings, this assumption frequently breaks down due to competing risks. A competing risk is an alternative event that is mutually exclusive with the primary outcome and physically prevents it from ever occurring. When standard survival models treat these competing events as simple censoring, they produce mathematically distorted, overly pessimistic estimates of risk. Competing risk analysis provides the statistical framework necessary to correct this distortion.
The Intuition: The Denominator Problem
To understand why standard methods fail, consider a study evaluating the long-term risk of developing a secondary bladder cancer following radiation therapy for rectal cancer, a scenario investigated by Guan et al..
Suppose a patient in the cohort dies from their primary rectal cancer two years into the study. Because they are dead, their risk of developing bladder cancer drops to absolute zero. If we analyze this dataset using a standard Kaplan-Meier curve, this patient's death is treated as a standard censored observation.
By censoring the patient, the statistical model removes them from the "risk set" (the denominator of individuals still capable of getting bladder cancer). However, the model simultaneously assumes that if we could have kept tracking this deceased patient, they would have continued to develop bladder cancer at the same rate as the surviving patients.
As more patients die from rectal cancer or other unrelated causes, the denominator artificially shrinks while the assumed risk remains unadjusted. This mathematical artifact causes the calculated cumulative incidence—the probability of the event occurring over time—to become severely overestimated. In aging populations or cohorts with high baseline mortality, this overestimation can lead clinicians to wildly inaccurate conclusions about treatment toxicities or disease progression.
Defining Competing Risks in Practice
A competing risk is not merely a confounding variable; it is an event that fundamentally alters the patient's trajectory. Across clinical research, this dynamic manifests in several distinct ways:
- Death vs. Non-Fatal Incidence: When studying the onset of dementia in heart failure patients, as seen in work by Ren et al., all-cause mortality acts as a competing risk. A patient who dies cannot subsequently be diagnosed with dementia.
- Intervention vs. Natural Progression: In a study of mortality risk factors for alcoholic hepatitis by Gasperment et al., receiving a life-saving liver transplant is a competing risk for death. The transplant abruptly removes the patient from the natural progression pathway of the disease.
- Surgical Revision vs. Mortality: When comparing total hip arthroplasty against hemiarthroplasty, Metcalfe et al. utilized competing risk frameworks because elderly patients frequently die from unrelated causes before their joint replacement has a chance to fail or require surgical revision.
Two Solutions: Cause-Specific vs. Subdistribution Hazards
To properly account for competing risks, researchers rely on two primary mathematical frameworks, each answering a fundamentally different clinical question.
┌───► Primary Event (e.g., Secondary Cancer)
│
Initial State ──────┼───► Competing Risk (e.g., Death from Primary Disease)
│
└───► Censoring (e.g., End of Study)
1. Cause-Specific Hazard Models
The cause-specific approach sets up separate Cox proportional hazard models for each distinct type of event. When analyzing the primary event, patients who experience the competing event are still treated as censored.
- The Intuition: This model calculates the event rate among those who are still alive and event-free at any given moment.
- When to use it: It is highly effective for studying etiology and biological mechanisms. For instance, Wang et al. used cause-specific Cox models to determine whether elevated glucose levels directly accelerate COVID-19 progression, treating hospital discharge as a competing risk.
However, cause-specific hazards do not directly translate into the actual probability that a patient will experience an event over a given time horizon, because they ignore the fact that the competing event is simultaneously removing people from the pool.
2. The Fine-Gray Subdistribution Hazard Model
To calculate true individual risk trajectories—known as the Cumulative Incidence Function (CIF)—researchers turn to the subdistribution hazard framework developed by Fine and Gray. This model was utilized by Patidar et al. to evaluate acute kidney injury outcomes while accounting for the competing risk of organ transplantation.
The core innovation of the Fine-Gray model lies in how it handles the risk set denominator. When a patient experiences a competing event (e.g., dies of an unrelated cause), they are intentionally left in the denominator of the risk set indefinitely, even though they can no longer experience the primary event.
- The Intuition: By keeping these individuals in the denominator with an event probability of zero, the model prevents the artificial inflation of the risk curve. The resulting hazard ratio represents the effect of a covariate on the actual likelihood of experiencing the primary event over time, accounting for the reality that some patients will be taken out of play by competing causes.
- When to use it: This is the gold standard for clinical prognosis, risk prediction algorithms, and medical decision-making, as it reflects the true real-world probability of an outcome for an individual patient.
Evaluating Differences: Gray's Test
Just as standard survival analysis relies on the log-rank test to compare two Kaplan-Meier curves, competing risk analysis requires a specialized metric to compare Cumulative Incidence Functions.
As highlighted by Drozd‐Sokołowska et al. in their study of lymphoma prognosis alongside comorbidities like diabetes, researchers use Gray's test. Unlike the log-rank test, which treats competing events as standard censoring and overestimates differences between groups, Gray's test mathematically accounts for the subdistribution hazards, ensuring that differences in cumulative incidence are truly driven by the primary disease process rather than imbalances in competing mortality rates.
By shifting from standard survival models to competing risk frameworks, researchers avoid statistical illusions, ensuring that risk scores, clinical guidelines, and epidemiological forecasts reflect genuine biological and clinical realities.
Key Foundational Work
Bayesian Survival Modelling of University Outcomes (2016) — contributed 1 concept elements to this direction.
Estimation of colorectal adenoma recurrence with dependent censoring (2009) — contributed 1 concept elements to this direction.