Flexible Parametric Survival Model (Royston-Parmar)
Also known as: flexible parametric model, restricted cubic spline survival model, stpm2, Esnek Parametrik Survival Modeli (Royston-Parmar)
The Royston-Parmar model, introduced by Royston and Parmar in 2002, is a modern parametric approach to survival analysis that replaces the rigid distributional assumptions of classical models with a restricted cubic spline fitted to the log-cumulative-hazard scale. It combines the interpretability of a fully parametric model with the flexibility to capture non-standard hazard shapes, and it supports proportional-hazards, accelerated failure-time, and proportional-odds link functions.
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 the Royston-Parmar model when you have right-censored time-to-event data with at least 50 subjects (to support the spline degrees of freedom) and when you suspect or observe that the hazard is not monotone or does not conform to a standard Weibull or exponential shape. It is particularly valuable when you need smooth absolute survival predictions, restricted mean survival time estimates, or hazard ratio estimates that are allowed to vary over time. It is also the preferred model when extrapolating survival beyond the observation period (health-technology assessment), because the parametric form produces well-behaved predictions in the tails. Covariates may be continuous or categorical. The model assumes that censoring is non-informative — subjects who leave follow-up carry the same subsequent risk as those who remain.
Strengths & limitations
- Captures non-monotone and non-standard hazard shapes without imposing a rigid distributional family.
- Produces smooth, analytically tractable survival and hazard functions, enabling direct computation of absolute risk differences and restricted mean survival time.
- Supports multiple link scales (PH, AFT, PO), making it more versatile than Weibull or Cox alone.
- Allows time-varying covariate effects (non-proportional hazards) by interacting predictors with the spline basis.
- Well suited to extrapolation in health-technology assessment because the parametric baseline behaves smoothly beyond the data range.
- Requires at least 50 events to support the spline; with fewer events the spline degrees of freedom should be kept very small (df = 1 reduces to Weibull) or a simpler model should be used.
- The choice of knot number and placement influences results; sensitivity analyses over alternative knot configurations are recommended.
- More complex to implement and interpret than Cox regression; not available in all mainstream statistical packages.
- Extrapolation beyond the data range depends heavily on the chosen knot structure and link scale, so clinical expert review is essential.
Frequently asked
How do I choose the number of knots?
A common starting rule is 1 knot (df = 2) for fewer than 100 events, 2 or 3 knots (df = 3–4) for 100–500 events, and up to 4 or 5 knots for larger datasets. Knots are typically placed at evenly-spaced centiles of the event-time distribution. Compare models by AIC across a small grid of df values and report a sensitivity analysis — results should be robust to adjacent knot configurations.
How does this model differ from Cox regression?
Cox regression leaves the baseline hazard completely unspecified (non-parametric), which is robust but means you cannot directly estimate the survival function at any time point without additional smoothing, and extrapolation is not meaningful. The Royston-Parmar model fully parameterises the baseline with a spline, so you get smooth absolute survival predictions, restricted mean survival time, and well-defined extrapolations — at the cost of needing to select knot configurations and link functions.
Can the model handle non-proportional hazards?
Yes. By interacting each covariate with the spline basis functions you obtain time-varying log-hazard-ratio estimates, allowing the hazard ratio to change smoothly over follow-up. This is a major advantage over the standard Cox model, where non-proportionality must be handled by stratification or adding explicit time-interaction terms.
What link function should I choose?
The proportional-hazards (PH) link is the default because its coefficients are directly comparable to Cox model hazard ratios. The accelerated failure-time (AFT) link gives time-ratio interpretations, which are sometimes more natural when the scientific question concerns how long survival is extended. The proportional-odds (PO) link converges to a log-logistic model. In practice, select the link that best fits the data (lowest AIC) and is most interpretable for the audience.
Sources
- Royston, P. & Parmar, M.K.B. (2002). Flexible Parametric Proportional-Hazards and Proportional-Odds Models for Censored Survival Data, with Application to Prognostic Modelling and Estimation of Treatment Effects. Statistics in Medicine, 21(15), 2175–2197. DOI: 10.1002/sim.1203 ↗
How to cite this page
ScholarGate. (2026, June 1). Flexible Parametric Survival Model (Royston-Parmar). ScholarGate. https://scholargate.app/en/survival/flexible-parametric-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.
- Accelerated Failure Time ModelSurvival↔ compare
- Bayesian Survival AnalysisBayesian↔ compare
- Cox RegressionSurvival↔ compare
- Fine-Gray Competing Risks ModelStatistics↔ compare
- Kaplan-MeierSurvival↔ compare
- Log-Rank TestSurvival↔ compare
- Mixture Cure ModelSurvival↔ compare
- Weibull RegressionSurvival↔ compare