Bayesian Survival Analysis
Also known as: bayesian sağkalım analizi, bayesian time-to-event analysis, bayesian hazard model
Bayesian survival analysis applies Bayesian inference to time-to-event models — Cox proportional hazards, parametric (Weibull, exponential), and cure models. Formalised comprehensively by Ibrahim, Chen and Sinha (2001), the approach encodes prior knowledge about hazard rates and regression coefficients, then updates it with censored survival data to yield posterior hazard ratios and credible intervals rather than single point estimates.
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
Bayesian survival analysis is the preferred approach when sample sizes are small (the registry sets a minimum of 20 observations), when events are rare and frequentist Cox estimates are unstable or underpowered, or when reliable prior information exists — for example from a previous study or a published meta-analysis — and should be formally incorporated. It is also valuable when the full posterior distribution of the hazard is needed rather than a single estimate, such as when downstream decisions depend on the probability that a hazard ratio exceeds a specific threshold. Noninformative censoring is required: the censoring mechanism must not depend on the underlying risk. A prior sensitivity analysis — comparing results under informative versus vague priors — is mandatory before reporting.
Strengths & limitations
- Encodes prior clinical or epidemiological knowledge about hazard rates and updates it formally with data.
- Produces posterior hazard ratios and credible intervals that are direct probability statements, valid even in small or rare-event samples where frequentist large-sample approximations are unreliable.
- Handles right-censored data exactly within the likelihood without imputation.
- Supports a wide class of baseline hazard specifications: nonparametric (Bayesian Cox), Weibull, exponential, and cure models.
- When n is small (close to the minimum of 20), the posterior is dominated by the prior; clinical interpretation requires caution.
- If prior sensitivity analysis reveals that the posterior shifts meaningfully with the prior choice, the Bayesian advantage over frequentist Cox is reduced.
- MCMC sampling is computationally heavier than frequentist maximisation, particularly for large datasets or complex baseline hazard specifications.
- Choosing and justifying the prior for the hazard requires domain knowledge; an uninformed choice can silently bias results.
Frequently asked
How does Bayesian survival analysis differ from frequentist Cox regression?
Frequentist Cox regression maximises the partial likelihood to obtain a single hazard ratio estimate with a standard error and a confidence interval based on large-sample theory. Bayesian survival analysis combines a prior distribution on the hazard parameters with the full survival likelihood to produce a posterior distribution for every quantity. Credible intervals are direct probability statements; they do not rely on asymptotic approximations and remain interpretable even with small samples or rare events.
What prior should I place on the hazard ratios?
Log-normal and Cauchy priors are the most common choices for log hazard ratios. A Cauchy(0, 2.5) prior — recommended by Gelman and colleagues for regression coefficients — places most mass near zero while allowing for large effects. The prior must always be justified and subjected to sensitivity analysis: run the model under both an informative prior (derived from previous studies) and a vague prior, and report both sets of results.
How do I check whether the MCMC has converged?
Run multiple chains (at least four) and compute R-hat for every parameter: values below 1.01 indicate convergence. Inspect trace plots for chain mixing — the chains should overlap and show no drift. Check the bulk and tail effective sample sizes, which should be at least a few hundred per parameter. Finally, perform posterior predictive checks by comparing simulated survival curves with the empirical Kaplan-Meier estimate.
What is noninformative censoring and why does it matter?
Noninformative censoring means that the reason a subject is censored — lost to follow-up, study ended, withdrew consent — is unrelated to their underlying risk of the event. If censoring is informative (e.g., sicker patients withdraw earlier), the survival likelihood no longer correctly represents the data-generating process and posterior estimates will be biased. Bayesian methods offer no special protection against informative censoring; the assumption must be verified by design and subject-matter knowledge.
Sources
- Ibrahim, J.G., Chen, M.-H. & Sinha, D. (2001). Bayesian Survival Analysis. Springer. DOI: 10.1007/978-1-4757-3447-8 ↗
How to cite this page
ScholarGate. (2026, June 1). Bayesian Survival Analysis. ScholarGate. https://scholargate.app/en/bayesian/bayesian-survival
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.
- Bayesian RegressionBayesian↔ compare
- Cox RegressionSurvival↔ compare
- Kaplan-MeierSurvival↔ compare
- Weibull RegressionSurvival↔ compare