Bayesian Competing Risks Analysis — Bayesian Competing Risks Survival Analysis
Bayesian Competing Risks Survival Analysis · Also known as: Bayesian cause-specific hazard model, Bayesian subdistribution hazard model, BCRA, Bayesian cumulative incidence analysis
Bayesian competing risks analysis is a time-to-event method for settings where subjects can fail from more than one mutually exclusive cause — such as death from cancer versus death from cardiovascular disease — and prior knowledge or small-sample uncertainty makes a Bayesian framework advantageous. It extends classical competing risks models (cause-specific hazards and cumulative incidence functions) by placing probability distributions over unknown parameters and updating those distributions with observed data, yielding full posterior inference for each failure type.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use Bayesian competing risks analysis when: (1) survival data contain two or more mutually exclusive failure causes and the interest lies in the probability of each cause or in covariate effects on cause-specific risks; (2) sample sizes are small or events are rare, making frequentist confidence intervals unreliable; (3) reliable prior information from previous studies or meta-analyses can be formally incorporated; or (4) full posterior uncertainty about the CIF or hazard ratios is required for decision-analytic modelling. Do NOT use it if you have only a single event type (standard Bayesian survival suffices), if event causes are not clearly distinguishable or independently verifiable, or if computational resources and MCMC expertise are unavailable — frequentist Fine–Gray or cause-specific Cox models are then the practical alternative.
Strengths & limitations
- Properly accounts for competing events, preventing the overestimation of cumulative incidence that arises from naive 1 − Kaplan–Meier estimation.
- Full posterior distributions enable direct probability statements about parameters and direct propagation of uncertainty into downstream cost-effectiveness or decision models.
- Informative priors allow borrowing strength from historical data, substantially improving estimates in small or sparse datasets.
- Simultaneous modelling of all competing causes maintains coherent estimation across causes and preserves the constraint that cause-specific CIFs sum to the overall incidence.
- Flexible baseline hazard specification (piecewise exponential, Weibull, non-parametric via Dirichlet process priors) accommodates varied hazard shapes.
- MCMC estimation is computationally intensive and requires convergence diagnostics; poorly specified models or diffuse high-dimensional priors can lead to slow mixing.
- Results can be sensitive to prior choice when data are sparse — an advantage when priors are well-justified but a vulnerability when they are not.
- Software implementation (Stan, JAGS, R packages such as BayesSurv or rstanarm extensions) requires greater statistical expertise than standard Cox or Fine–Gray regression.
- Requires clear, reliable cause-of-death or cause-of-event ascertainment; misclassified competing events propagate bias that Bayesian machinery does not correct.
Frequently asked
What is the difference between cause-specific hazard and subdistribution hazard in this context?
The cause-specific hazard for event k is the instantaneous rate of event k among those still event-free. The subdistribution hazard (Fine–Gray) for event k is defined over a risk set that retains subjects who have already experienced a competing event. Both can be parameterised in a Bayesian framework, but they answer different questions: cause-specific hazards describe the biological or mechanistic rate; subdistribution hazards directly predict cumulative incidence and are preferred for prognosis. Bayesian competing risks analysis can implement either or both.
Why not simply censor competing events and use a standard Bayesian survival model?
Censoring competing events treats them as uninformative, which is almost always false — a patient who dies from another cause is systematically different from one who is administratively censored. This inflates the cause-specific cumulative incidence and produces biased predictions. Competing risks analysis avoids this by explicitly modelling all failure types simultaneously.
How do I choose priors for the regression coefficients?
A common starting point is weakly informative priors on log hazard ratios, such as Normal(0, 1) on the log scale, which shrink implausibly large effects toward zero. When historical evidence is available — for instance, from a meta-analysis reporting a log hazard ratio and its standard error — a normal prior centred on that estimate is appropriate. Always perform a prior sensitivity analysis by repeating the analysis with both more and less informative priors to assess how much the posterior depends on prior choice.
Which software packages support Bayesian competing risks models?
Stan (via RStan or CmdStanR) allows full custom specification. The R packages rstanarm and brms do not directly implement competing risks but can be extended. Dedicated packages include BayesSurv and CompRisks in R. JAGS and BUGS can also be used with user-written model code. For straightforward piecewise exponential competing risks models, a Stan template requires fewer than 100 lines of code and is often the most transparent approach.
Is Bayesian competing risks analysis suitable for randomised clinical trials?
Yes, and it is increasingly used in small or adaptive trials where frequentist power is limited. In a randomised setting the Bayesian approach allows incorporation of prior data from earlier phases, direct probability statements such as the probability that treatment reduces cause-specific incidence by more than a clinically meaningful threshold, and seamless updating as interim data accumulate. The causal interpretation of hazard ratios requires the same assumptions as in frequentist analysis.
Sources
- Larson, M. G., & Dinse, G. E. (1985). A mixture model for the regression analysis of competing risks data. Applied Statistics, 34(3), 201–211. DOI: 10.2307/2347464 ↗
- Crowder, M. J. (2001). Classical Competing Risks. Chapman and Hall/CRC. ISBN: 9781584881759
How to cite this page
ScholarGate. (2026, June 3). Bayesian Competing Risks Survival Analysis. ScholarGate. https://scholargate.app/en/epidemiology/bayesian-competing-risks-analysis
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Cox proportional hazardsEpidemiology↔ compare
- Kaplan-Meier EstimatorStatistics↔ compare
- Multi-State ModelSurvival↔ compare