Statistics Glossary → Competing Risks
Competing Risksanalysis
A survival analysis for situations with several event types, where the occurrence of one makes the others impossible, that estimates the cumulative probability of each event over time and the effects of factors on these events.
Concept
In survival data there are often other events, besides the one of interest, that prevent it: when studying cancer relapse, some patients die of another cause before relapsing; when studying resignation, some employees retire. A person who has this 'competing' event can no longer have the event of interest. Treating competing events as ordinary censoring and computing the event probability as 1 - Kaplan-Meier assumes these people could still have the event later, and so overstates the probability.
The key quantity in competing risks analysis is the cumulative incidence function (CIF): the probability that a given event type has occurred by a given time, estimated with the Aalen-Johansen method while allowing for the competing events. The effects of factors are studied with two different models. The cause-specific Cox model gives the effect on the instantaneous rate of that event among people who have not yet had any event (cause-specific HR) and answers 'does the factor change how fast this event occurs?'. The Fine-Gray model gives the effect of the factor on the cumulative incidence of that event (subdistribution HR, sHR) and answers 'does the factor change the share of people who have this event?'.
Because the two models answer different questions, their results can differ, even in direction: if a factor greatly speeds up the competing event, it lowers the cumulative incidence of the event of interest even with no effect on its rate. The two models are therefore often reported together. Differences in CIF between groups are tested with Gray's test. In the cause-specific models the proportional hazards assumption should be checked, and censoring is assumed to be non-informative.
When to use it
- When other events prevent the event of interest and the question is 'what share of people have this event within a given time?' (relapse and death before relapse, graft loss and death with a functioning graft, resignation and retirement).
- If the question is about a factor's biological or causal effect on how fast the event occurs, the cause-specific Cox model comes first; if risk prediction and the share who will have the event matter, the Fine-Gray model does.
- If there is no competing event or it is very rare, Kaplan-Meier and Cox regression are enough; if the events do not preclude each other but recur in the same person, recurrent-event models are used.
An everyday example
A company follows 1,000 employees for 10 years to see how long it takes them to resign. Some employees retire without resigning, and someone who has retired can no longer resign. 1 - Kaplan-Meier might put the share resigning within 10 years at 45%, while the cumulative incidence that treats retirement as a competing event might find 35%. The cause-specific HR for age may be close to 1 for resignation; but because older employees retire sooner, the Fine-Gray sHR is below 1: the share who resign is lower among older employees.
Parameters in MerQur
Time column, Event type (0 = censored) and Censoring value
Concept: Time is the period from the start of follow-up to the first event or to censoring. The event-type column shows whether each person is censored (0) or which event they had (1, 2, …); if censoring is coded with another value, that value is stated.
Everyday example: In the company example, time is the number of years from hiring to leaving or to the end of the study. Event type is coded 0 = still employed, 1 = resignation, 2 = retirement.
What it does: The event type column is coded 0 = censored and 1..K as causes; if a censoring value is written, that value is counted as 0 and the other values as sequential 1..K.
Effect on the result: The cause mapping and all counts change.
On the validation data: Melanoma, censoring value "2": the code mapping is 1 → 1 (melanoma), 3 → 2 (other); event counts 134 / 57 / 14. When left empty, a column that does not contain 0 gives the error "No censored observation (0)".
Event 1 label, Event 2 label
Concept: Gives readable names to the event codes; these names appear in the result tables instead of '1' and '2'. It does not affect the calculations.
Everyday example: Entering 'Resignation' for 1 and 'Retirement' for 2 lets the tables and plots be read directly with these names.
What it does: The event name that appears in the output.
Effect on the result: Only the text label changes; the numerical result does not change. With text-coded or nonzero censoring, the default labels do not overwrite the original values.
On the validation data: The numbers of the labeled and unlabeled runs are the same.
Covariates
Concept: These are the explanatory variables in the cause-specific Cox and Fine-Gray models. Each covariate's effect is estimated separately for each event type; the same covariate may raise the rate of one event while leaving the other unchanged.
Everyday example: Age, salary level and department are added as covariates. Salary may lower the resignation rate but not affect retirement, while age raises the retirement rate sharply.
What it does: The covariates of the cause-specific Cox and Fine-Gray models. If empty, only the event counts, the partial LL of the empty model and the cumulative incidence are given.
Effect on the result: The coefficients change.
On the validation data: Bmt Disease (baseline level ALL): Fine-Gray -0.803372 (SE 0.428456) and 0.508515 (0.366182). mgus2 Fine-Gray sexM (PCM) -0.229237.
Categorical (factor) variables and Reference levels
Concept: Lets a numerically coded categorical variable be modelled with dummy variables; the reference level sets the group the other levels are compared with.
Everyday example: If department is coded 1 = production, 2 = sales, 3 = management, it is made a factor with 'production' as reference. The HRs for sales and management are read relative to production.
What it does: Makes a numerically coded column a factor; the reference level field sets the reference of a text covariate (if a numeric column name is written, it makes the column categorical).
Effect on the result: The dummy variables and the reference of the coefficients change.
On the validation data: Bmt Disease with baseline level "Disease=1" gives 2 dummies; when empty, Disease is a single numeric term (Fine-Gray 0.369358); with baseline level "Disease=3", Disease_1 -0.508515, Disease_2 -1.311887.
CIF group variable (optional)
Concept: Separate cumulative incidence curves are drawn for each group of the chosen variable, and Gray's test checks whether an event's cumulative incidence differs between groups. Gray's test is the competing-risks counterpart of the log-rank test in Kaplan-Meier.
Everyday example: Grouping by department draws a 'share resigning within 10 years' curve for each department. If Gray's test is significant, the resignation rate in the sales department really differs from the others.
What it does: Cumulative incidence per group (n ≥ 10) and the Gray test.
Effect on the result: Only the group-based incidence and test change.
On the validation data: Bmt Disease: Gray χ² = 11.9229 (2 df, p = 0.002576). Melanoma ulcer χ² = 26.1207; mgus2 sex death χ² = 11.6513.
CIF time points (optional)
Concept: Sets the times at which the cumulative incidence and its standard error are reported, so that 'the share who had the event within a given time' can be read directly.
Everyday example: If years 2, 5 and 10 are chosen, the table might show 12%, 24% and 35% for resignation and 3%, 10% and 22% for retirement.
What it does: The times at which the overall cumulative incidence and the SE are reported.
Effect on the result: Only the reported times change.
On the validation data: Bmt (years) relapse CIF(0.5) = 0.145985 (SE 0.030296), CIF(2) = 0.301199 (0.039566); the SE is computed with the Gray variance.
Tie handling (cause-specific Cox) (Efron / Breslow)
Concept: Sets how the partial likelihood of the cause-specific Cox model is computed when several events happen at the same time (ties). The Efron approximation is usually more accurate; with few ties, the two methods give almost the same result.
Everyday example: If leaving dates are recorded only by month, several resignations fall in the same month. The Efron method handles these ties better than Breslow; with 1-2 ties per month the difference is negligible.
What it does: The ties approximation for the cause-specific Cox model; Fine-Gray always uses Breslow (crr).
Effect on the result: The cause-specific Cox coefficients change.
On the validation data: Bmt Disease_3 relapse: Efron 0.6107, Breslow 0.6101.
Additional output: CIF confidence intervals (log-log; table + plot bands), Proportional hazards test (cause-specific Cox, Schoenfeld), Omnibus tests (LR / Wald / score), Number-at-risk table
Concept: Confidence intervals for the cumulative incidence show its uncertainty; the test based on Schoenfeld residuals checks the proportional hazards assumption in the cause-specific Cox model; the omnibus tests (LR, Wald, score) say whether the covariates are jointly significant; the number-at-risk table shows how many people are still followed at given times.
Everyday example: If salary's effect on resignation is strong in the early years and then fades, the proportional hazards test is significant. If the number-at-risk table shows only 80 employees left at year 10, estimates for that period are uncertain.
What it does: The log(-log) interval for CIF; for the cause-specific Cox, the Grambsch-Therneau PH test with km transform; the LR, Wald and score omnibus tests; the number at risk at given times, cumulative event counts.
Effect on the result: Only adds output.
On the validation data: Bmt CIF(0.5) interval [0.0928, 0.2106] (log-log transformed interval); relapse Disease PH χ² = 4.0488 (2 df, at the term level); relapse omnibus LR 15.8886, Wald 14.3252, score 16.4930 (Efron).
Confidence Level (%90 / %95 / %99 / Other (custom))
Concept: Sets the confidence level of the interval estimates for the subdistribution hazard ratios (sHR) and cumulative incidence; if an sHR's interval excludes 1, the factor's effect on cumulative incidence is significant.
Everyday example: If the sHR for age is 0.97 for resignation with a 95% confidence interval of 0.95-0.99, the share who resign falls clearly with age; choosing 99% widens the interval.
What it does: The level of the subdistribution HR (sHR) and cumulative incidence intervals; the significance mark is given according to the selected α.
Effect on the result: Only the intervals change.
On the validation data: Bmt Fine-Gray Disease_2 sHR interval: 95% (0.1934, 1.0371), 90% (0.2213, 0.9061), 99% (0.1485, 1.3502).
Example on real data
Data: MGUS cohort, R survival mgus2
Variables: time etime (months), event type event (1 = plasma cell malignancy, 2 = death), covariates age, sex, mspike
- Question: In MGUS patients (n = 1373), plasma cell malignancy (PCM, 115 events) and death without PCM (854 events) are events that prevent each other; how do age, sex and the M-protein level (mspike) affect the risk of these events?
- Cumulative incidence: The final Aalen-Johansen CIF is 0.1619 for PCM and 0.8381 for death; about one in six patients develops PCM by the end of follow-up.
- Cause-specific Cox: mspike increases the PCM hazard (HR = 2.42 [1.751, 3.346], p < .001), age is borderline (HR = 1.0165, p = 0.051). Age (HR = 1.0674, p < .001) and male sex (HR = 1.475, p < .001) increase the hazard of death.
- Fine-Gray: For the cumulative incidence of PCM, mspike sHR = 2.43 [1.794, 3.296] (p < .001); age sHR = 0.983 (p = 0.0037), that is, older patients reach PCM less often because of the competing death. The cause-specific age effect is positive while the Fine-Gray effect is negative: the two models answer different questions.
- Limitation: Single-event Cox or 1 - KM overstates the PCM probability when competing risks are present; the cause-specific HR should be interpreted for etiology, the sHR for risk prediction.
- APA: In the Fine-Gray model, mspike, sHR = 2.43, 95% CI [1.79, 3.30], p < .001, and age, sHR = 0.98, 95% CI [0.97, 0.995], p = .004, significantly predicted the cumulative incidence of PCM.
Related concepts
Censoring (survival analysis)Ties method (Efron, Breslow, exact)Cumulative incidence function (CIF) and Gray's testFine-Gray subdistribution hazard modelCause-specific hazardCovariateKaplan-Meier estimator (survival curve)Dummy (indicator) variable coding
The numbers in the examples are MerQur's output on independent, publicly available validation data; see the MerQur Technical Note for details.