Parametric Survival Model

Parametric Survival Model: Concept Architecture

Orientation and learning roadmap

This page follows a time-to-event question from censored observations to a fitted survival curve and its long-term projection. It uses an exponential example to connect survival, hazard, and cycle probabilities, then sets out the validation needed before those projections enter an economic model.

Define the event and follow-up

Survival analysis needs a clearly defined event, time origin, follow-up window, and population at risk. A person whose event has not occurred by the last observed time may be right-censored rather than event-free forever; the likelihood uses the information available up to censoring under the model's assumptions.

Let $T$ be event time, $S(t)=P(T>t)$ the survival function, $h(t)$ its hazard when defined, and $H(t)$ the cumulative hazard. For a continuous nonnegative event time under the usual regularity conditions, $S(t)=\exp[-H(t)]$ and $h(t)=-d\log S(t)/dt$. A parametric model specifies a distribution or functional form for $T$ conditional on covariates, allowing both interpolation within observed follow-up and extrapolation beyond it.

Choose a shape that matches the question

Different distributional families imply different hazard shapes and therefore different long-term trajectories. The choice should be assessed against clinical knowledge and data, not selected by a fit statistic alone when the decision horizon extends far beyond the observed period.

Treatment effects can be specified through proportional hazards, accelerated failure time, separate treatment-arm curves, or more flexible time-varying relationships. Those parameterisations do not imply identical interpretations of a coefficient. When a constant hazard ratio is implausible, enforcing proportional hazards can misstate survival differences and downstream quality-adjusted life-years (QALYs).

Worked example: an exponential model

Suppose an illustrative event-time model has a constant hazard $\lambda=0.20$ per year for a specified population. This is a teaching assumption, not a mortality estimate from a real trial; it implies the survival and cumulative hazard functions below.

$$S(t)=e^{-0.20t},\qquad H(t)=0.20t,\qquad h(t)=0.20\ \text{per year}.$$

At two years, the event-free probability is $S(2)=e^{-0.40}\approx0.6703$. Conditional on surviving to the beginning of any one-year interval in this constant-hazard model, the probability of an event during that interval is $1-e^{-0.20}\approx0.1813$. The latter is a probability over one year, whereas $0.20$ has units of inverse years and is not itself the annual event probability.

Translate a curve into model cycles

A state-transition model may need a conditional event probability for a cycle from $t$ to $t+\Delta$, rather than survival from time zero. With one specified event and the corresponding survival function, this probability follows from the ratio of endpoint survival values. The conversion is shown below.

$$q(t,t+\Delta)=1-\frac{S(t+\Delta)}{S(t)},\qquad S(t)>0.$$

The formula assumes that $S$ describes the event being translated and that conditioning on event-free survival is meaningful for the state at cycle start. Competing events or multiple health states require an appropriate competing-risk or multistate construction; independently converting separate cause-specific curves into transition probabilities can make totals inconsistent. QALYs and life-years also depend on the area under the relevant survival or state-occupancy curves, not just the value at one endpoint.

Fit, inspect, and validate extrapolation

Fitting uses observed event times and censoring information, often by likelihood methods with covariates when needed. A fitted curve should be compared with Kaplan–Meier survival estimates within follow-up, observed and predicted hazards, numbers at risk, and clinically relevant subgroup patterns.

Information criteria such as AIC or BIC can help compare models on the observed data, but near-identical in-sample fit may coexist with very different long-run tails. External cohorts, background mortality, disease knowledge, and clinician judgement can inform the plausibility of projections, with discrepancies reported rather than concealed. Uncertainty analysis should vary both estimated parameters and materially credible structural alternatives.

What the economic decision needs

Survival extrapolation can dominate lifetime costs, life-years, and QALYs when trial follow-up is short relative to the decision horizon. The selected model should make clear whether survival is overall survival, progression-free survival, time to treatment discontinuation, or another endpoint; those curves cannot simply be substituted for one another.

For a treatment comparison, avoid implausible long-term separation or crossing by checking the clinical mechanism and external evidence, while allowing such patterns when substantively supported. If treatment effects wane or background mortality changes with age, a simple constant-effect extrapolation may need adjustment. Report the model family, parameterisation, fitting population, time units, censoring assumptions, tail behaviour, and impact of alternative models on incremental net benefit.

Sources and further learning

The linked technical guidance addresses choosing and validating survival extrapolations for economic evaluations. The methodological paper extends familiar parametric families with flexible hazard and odds structures, illustrating why assumptions about the curve shape matter.

• NICE Decision Support Unit, Technical Support Document 14, Survival Analysis for Economic Evaluations Alongside Clinical Trials: Extrapolation with Patient-Level Data. This covers parametric survival fits and long-term validation.
• NICE, NICE technology appraisal and highly specialised technologies guidance: the manual, economic evaluation. This requests survival curves, hazard plots, and support for extrapolation.
• Royston and Parmar (2002), “Flexible Parametric Proportional-Hazards and Proportional-Odds Models for Censored Survival Data”, Statistics in Medicine. This develops flexible time-to-event models.