Parametric survival & extrapolation
Fit a survival distribution and explore survival beyond the observed follow-up period.
Concepts and reasoning
Specify a shape for the whole survival curve
A parametric survival model gives the event-time distribution a mathematical form. That form determines survival probabilities, hazards, quantiles, and often the predicted mean survival. It can provide smooth estimates and extrapolate beyond the observed follow-up, which is useful when decisions require longer horizons than a study observes.
The benefit and the risk come from the same source: the distribution supplies a tail where there are little or no data. Two models can agree closely during the observed period and diverge sharply afterward. An attractive fit to early observations does not establish which extrapolation is credible.
Learn what each shape allows
An exponential model has a constant hazard. A Weibull model allows a hazard that increases or decreases monotonically. Lognormal and log-logistic models can describe hazards that rise and then fall. These are qualitative restrictions on how risk changes with time, not merely different labels on a software menu.
A model should be chosen with the clinical process, endpoint, follow-up pattern, and external evidence in mind. A sudden treatment-effect waning, a cure fraction, or population background mortality may not be represented adequately by a single simple distribution. The lifelines regression guide explains parametric and accelerated-failure-time formulations.
Mathematical foundations
Connect survival to hazard
If
For an exponential model with rate
A Weibull model can be written with scale
When
Censored observations still affect estimation
An observed event at time
Worked example
Equal medians do not imply equal tails
Consider two hypothetical models with median survival 12 months. For the exponential model,
Now use a Weibull model with shape 2 and choose scale so the same 12-month median is retained. The scale is
| Time in months | Exponential survival | Weibull survival, shape 2 |
|---|---|---|
| 12 | 0.500 | 0.500 |
| 24 | 0.250 | 0.063 |
| 36 | 0.125 | 0.002 |
The two models share a median but encode very different subsequent hazards. A decision model using years of extrapolated survival would obtain different health and cost consequences from these tails.
If only a short period around the median is well observed, the data may not distinguish these long-term predictions strongly. The next step is not to choose the more favorable tail. Examine clinical plausibility, later external observations, and uncertainty across defensible model structures.
Interpreting the results
Separate interpolation from extrapolation
Within observed follow-up, the fitted survival function smooths the available event and censoring information. Beyond the data, its shape is primarily a consequence of the chosen distribution. Two models can fit the observed period similarly yet imply very different long-term survival.
Akaike and Bayesian information criteria compare fit with a penalty for model complexity on the same data. Lower values support relative in-sample fit; they do not show that an extrapolated tail is biologically credible. Distribution selection remains a substantive modeling decision.
A survival probability at time t is the estimated fraction event-free beyond t under the model. Pointwise confidence intervals quantify parameter uncertainty conditional on that distribution; they are not simultaneous uncertainty bands for every time and do not capture uncertainty about the family itself. Label extrapolated values clearly when communicating results.
Assumptions and limitations
Distributional shape is a strong assumption
Exponential survival implies a constant hazard. Weibull survival allows a monotone hazard, while lognormal and loglogistic families allow different nonmonotone patterns and tails. Choose among them using event mechanisms, observation length, and external evidence as well as fit statistics.
The usual likelihood assumes independent observations and noninformative right censoring. Different populations, treatment switching, competing events, or long-term background mortality may require additional structure. A pooled unadjusted fit describes the supplied mixture and should not be treated as a covariate-adjusted treatment comparison.
The time scale, starting point, and endpoint must be consistent. Long survival tails can imply large accumulated life-years even when few participants remain observed. The apparent smoothness of a fitted curve does not indicate that those late predictions are supported by data.
Diagnostics and sensitivity analysis
Compare the fitted period and the tail separately
Overlay candidate curves with an empirical survival estimate and inspect regions with many observations as well as the sparse tail. Examine parameter estimates and convergence warnings; a failed fit is evidence to investigate, not merely an inconvenient missing curve.
Repeat extrapolation using plausible alternative distributions and horizons. Describe how much the decision depends on survival outside observed follow-up. Compare predictions with external follow-up or clinical expectations where available, without treating reused calibration data as independent validation.
Inspect implied hazards and long-term behavior, not only survival probabilities at a single date. If several plausible families disagree materially, report that structural uncertainty. The lifelines univariate modeling guide explains the supported distributional forms and estimation tools.
References and other material
Read the methodological references and supporting tutorials below. The examples in this lesson are original teaching calculations, not findings from these sources.
- lifelines: survival regression — Worked regression examples and assumption checks.