Flexible Parametric Survival Models (Royston–Parmar) – Solving Extrapolation

Research updated on January 11, 2026
Author: Santhosh Ramaraj

Let’s start with the simplest possible intuition.

You are trying to understand how long people live after starting a treatment. You don’t just want to know who does better, you want to know how the survival curve behaves over time, especially after your data ends. That last part is the key reason this model exists.

In HEOR, survival analysis is not only about statistics. It directly determines life-years, QALYs, costs, and ICERs. If your survival curve is wrong, everything downstream is wrong too.

Flexible parametric survival models, often called Royston–Parmar models, were designed specifically to solve the problem of realistic, transparent survival extrapolation.

Why extrapolation is unavoidable

Imagine you run a study for 3 years. Some patients are still alive at the end. But a cost-effectiveness model might need survival estimates over 20, 30, or even 50 years.

You are forced to answer questions like:

What happens after year 3?
Does survival slowly decline?
Does it flatten?
Does treatment benefit disappear?

You cannot avoid making assumptions. The only question is whether your assumptions are rigid and hidden or flexible and explicit.

Traditional parametric models like exponential or Weibull force survival into very specific shapes. Sometimes that’s fine. Often, especially in oncology, it’s not.

Royston–Parmar models were created to let the data guide the shape, while still producing smooth curves that behave sensibly in the long term.

What makes Royston–Parmar model “flexible”

Instead of assuming a fixed hazard shape, the Royston–Parmar model uses splines to model how risk changes over time.

A spline is just a smooth curve that can bend. Think of time on the x-axis and risk on the y-axis. Instead of drawing one simple curve, you allow the curve to bend gently at certain time points. These bend points are called knots.

The important thing is that the curve stays smooth. There are no sharp corners or jumps. This matters because survival curves must look biologically plausible.

Mathematically, the model often works on the log cumulative hazard scale. In simple terms, instead of modeling survival directly, it models a transformed version that is easier to control and smooth.

A basic form looks like this:

$\log H(t) = \beta_0 + \beta_1 s_1(t) + \beta_2 s_2(t) + \dots + \beta_k s_k(t)$

Here,
$H(t)$ is the cumulative hazard at time, $t$ and the spline terms $s_k(t)$ are smooth functions of time. Each spline basis function contributes to shaping the overall hazard curve.

You don’t need to calculate these by hand. What matters is understanding that each term adds controlled flexibility, not randomness.

How treatment effects are included

Now suppose you are comparing two treatments: standard care and a new drug.

You add treatment as a covariate, just like in a Cox model:

$\log H(t) = \eta(t) + \gamma \cdot \text{Treatment}$

This already gives you separate survival curves for each treatment. But here’s where Royston–Parmar goes beyond Cox.

You can allow the treatment effect itself to change over time.

This is extremely important in HEOR, especially for immuno-oncology, where treatment benefit is often strong early and weaker later.

You can write:

$\log H(t) = \eta(t) + \gamma(t) \cdot \text{Treatment}$

Now $\gamma(t)$ is also a spline. That means the treatment effect can fade, grow, or stabilize over time, based on the data.

Cox models assume proportional hazards unless you force complicated fixes. Royston–Parmar models handle this naturally.

A example you can picture

Let’s say we simulate data for 600 cancer patients. Half receive standard chemotherapy. Half receive a new targeted therapy.

We follow them for 36 months. During that time, the new therapy clearly improves survival. After 36 months, many patients are still alive, and we have no direct data beyond that.

We fit a Royston–Parmar model with three internal knots. The model learns:

Early hazard is lower for the new drug.
Mid-term hazard slowly increases.
Long-term hazard flattens.

The resulting survival curve looks smooth and believable. When we extrapolate beyond 36 months, the curve does not suddenly crash or unrealistically plateau. It follows the logic implied by earlier trends.

This is exactly what HTA reviewers want to see. Not perfection, but transparent, defensible assumptions.

Why this matters for economic models

Once you have a survival curve, everything else flows from it.

Expected survival is calculated as the area under the survival curve:

$\text{Mean Survival} = \int_0^T S(t),dt$

That number feeds directly into QALYs and ICERs.

For example:

$ICER = \frac{\mu_{C1} - \mu_{C2}}{\mu_{E1} - \mu_{E2}}$

If your extrapolated survival curve is too optimistic, the ICER looks better than it should. If it’s too pessimistic, you might reject an effective treatment. This is why HEOR teams prefer flexible parametric models. They reduce the risk that model choice, not evidence, drives decisions.

Choosing the number of knots (intuitively)

You don’t want the curve too stiff, and you don’t want it too wiggly. In practice:

Most HEOR analyses use 2 to 4 knots.
Fewer knots give smoother, more conservative extrapolation.
More knots fit observed data better but risk unstable tails.

The important thing is not the exact number, but that you test sensitivity and justify your choice clinically.

Extrapolation is an assumption, not a fact

This is something many people misunderstand. Royston–Parmar models do not magically predict the future. What they do is allow you to clearly show how your assumptions shape the future.

You can:

Compare spline models to Weibull or exponential tails.
Test optimistic vs pessimistic scenarios.
Align long-term behavior with external registry data.

That transparency is why agencies like NICE accept these models when they are well justified.

Final intuition to keep in your head

Think of the Royston–Parmar model as a smart curve-drawing tool.

It lets you draw survival curves that: Follow the data closely where data exists, behave sensibly where data does not exist, and can be explained, defended, and tested.

Relevant Reads