Parametric Survival Analysis: Modelling Time-to-Event Data When the Hazard Has a Predictable Shape

Mayta
Jun 17
3 min read

Introduction

Time-to-event studies frequently rely on Cox proportional-hazards regression because it requires no a priori specification of how risk evolves. Yet there are many real-world situations—early postoperative deaths, age-related mortality, tumour-recurrence peaks—where the hazard follows a pattern that can be described by a mathematical distribution. Parametric survival analysis exploits that knowledge: by choosing a distribution whose parameters control the shape of the hazard, investigators can estimate absolute survival probabilities, mean survival times, and other quantities that the semi-parametric Cox model leaves undefined.

1 Why Specify a Distribution?

Instantaneous risk is more informative than its average. When the hazard rises, falls, or re-emerges, a single summary measure may conceal clinically important phases of risk.
Parameters map directly onto hazard shape. Selecting a distribution fixes how risk changes over time; its parameters then quantify how quickly those changes occur.
Richer predictions are possible. Once the hazard is fully defined, one can calculate not only hazard ratios but also survival curves, event-free life expectancy, and restricted mean survival times for any covariate pattern.

Investigators usually justify a chosen distribution through prior knowledge, empirical inspection of pilot data, or theoretical considerations about pathophysiology.

2 Common Distributions and Their Hazard Shapes

Distribution	Key Parameters	Typical Hazard Shape	Clinical Illustration*
Exponential	single rate (λ)	Constant over time	Radioactive tracer decay
Weibull	scale (λ), shape (ρ)	Monotonically ↑ if ρ>1, ↓ if ρ<1	Device failure accelerating with wear
Gompertz	baseline (λ), growth (γ)	Exponentially increasing	Age-related all-cause mortality
Log-logistic	scale (λ), shape (γ)	Inverted U-shape (rises then falls)	Post-operative infection risk peaking at day 7
Log-normal	mean (μ), SD (σ) of log time	Inverted U-shape, longer tail	Incubation period of viral disease
Gamma	scale (λ), shape (κ, ρ)	U-shape or unimodal, very flexible	Competing biological repair and damage processes

*Examples are illustrative substitutes, not taken from the slide set.

A concise reference table on page 4 of the source summarises each distribution’s parameter count and monotonicity characteristics.

3 Visualising How Distributions Differ

Consider a graph in which each coloured curve depicts how the hazard evolves over twelve years for the distributions above. Some curves climb steadily, others peak early then decline, and one remains flat. Such visualisation clarifies why an exponential model would be inappropriate if early risk spikes are expected, or why a log-logistic model handles hazards that eventually wane.

4 Fitting Parametric Models in Practice

4.1 Model Estimation

Most statistical packages offer a single command that combines:

Declaring survival data (e.g., stset in Stata),
Choosing a distribution (e.g., streg , dist(weibull)),
Specifying covariates exactly as in Cox regression.

The maximum-likelihood routine estimates both regression coefficients and distribution parameters simultaneously. Output includes hazard ratios that resemble those from Cox models, plus baseline hazard or survival functions.

4.2 Generating Predictions

With parameters in hand, researchers can:

Plot model-based survival curves for contrasting covariate patterns (stcurve, survival at() in Stata).
Compare curves to Kaplan–Meier estimates to verify that the chosen distribution reproduces observed data adequately. An example workflow on the slide set overlays modelled curves onto non-parametric estimates to check goodness-of-fit.

4.3 Model Adequacy Checks

Graphical overlay: Do model-based curves track empirical Kaplan–Meier steps?
Residuals or information criteria: Compare alternative distributions through AIC/BIC.
Clinical plausibility: Does the hazard shape align with disease biology?

5 When Parametric Models Outperform Cox

Question	Cox Model	Fixed-Shape Parametric Model
Baseline hazard explicitly estimated?	✘	✔
Absolute survival at 1, 5, 10 years?	⚠ (needs baseline); indirect	✔ direct
Mean survival time (restricted or total)?	Requires numerical integration of baseline	✔ closed-form
Extrapolation beyond follow-up?	Hazard unknown	✔ if distribution justified
Flexibility if shape is unknown?	Good (semi-parametric)	Potentially poor

A summary matrix on the final slide emphasises that Cox’s convenience-no baseline hazard assumption—comes at the cost of absolute predictions. Parametric models reverse that trade-off.

6 Beyond Fixed-Shape Models

Flexible parametric models introduce restricted cubic splines within the hazard, blending flexibility with parametric prediction.
Accelerated failure-time (AFT) models express covariate effects on survival time rather than on hazard.
Laplace regression and restricted mean survival time (RMST) models target survival percentiles or average event-free time directly.

These variants preserve the predictive advantages of parametric analysis while relaxing rigid shape constraints.

Conclusion

Parametric survival analysis supplies a powerful toolkit for studies where the course of risk is biologically or empirically predictable. By selecting a distribution whose parameters capture that pattern, investigators gain access to absolute survival estimates, life-expectancy calculations, and credible extrapolation—outputs unattainable with semi-parametric methods alone. The choice of distribution must, however, rest on sound evidence; a mis-specified hazard can obscure true effects just as surely as an inappropriate linearity assumption in simple regression. When thoughtfully applied, parametric models illuminate the full trajectory of risk and enrich the interpretation of longitudinal clinical data.