Skip to main content

Parametric Survival Models

Modeling event times with an explicit distribution

Statistics
Advanced Modeling
Survival Analysis
Parametric survival distributions, proportional-hazards and accelerated-failure-time interpretations, estimation, diagnostics, and design-aware bioinformatics use.
Author

Jiahong Sun

Published

November 7, 2026

Modified

November 7, 2026

1 Put a Distribution on Event Time

A parametric survival model specifies the entire conditional distribution of an event time \(T\) rather than leaving the baseline hazard unspecified. For covariates \(x\), the model determines mutually linked quantities

\[ S(t\mid x)=P(T>t\mid x),\qquad f(t\mid x)=-\frac{d}{dt}S(t\mid x), \]

\[ h(t\mid x)=\frac{f(t\mid x)}{S(t\mid x)},\qquad H(t\mid x)=-\log S(t\mid x). \]

This extra structure can estimate absolute risks, survival quantiles, restricted mean survival times, and extrapolated survival beyond the observed follow-up. It also creates extra ways to be wrong. A smooth fitted curve is evidence that a distribution was fitted, not that its tail or extrapolation is credible.

The experimental unit is the independently sampled subject or biological unit whose event history is modeled. Repeated assays, tissue sections, cells, or sequencing reads from one subject can improve covariate measurement but do not create additional independent survival times.

2 Common Baseline Distributions

Suppress covariates temporarily. A useful distribution is characterized not only by its density but by the hazard shapes it permits.

Distribution Survival or hazard Hazard shape Important implication
Exponential, rate \(\lambda>0\) \(S(t)=e^{-\lambda t}\); \(h(t)=\lambda\) Constant Memoryless; often too restrictive
Weibull, scale \(a>0\), shape \(p>0\) \(S(t)=\exp\{-(t/a)^p\}\) Decreasing if \(p<1\), constant if \(p=1\), increasing if \(p>1\) Supports both PH and AFT forms
Gompertz, parameters \(b>0\), \(c\ge 0\) \(h(t)=b e^{ct}\) Constant or exponentially increasing Useful for approximately exponential age effects on mortality
Log-normal, \(\log T\sim N(\mu,\sigma^2)\) \(S(t)=1-\Phi\{(\log t-\mu)/\sigma\}\) Rises then falls Natural AFT model; not PH
Log-logistic, \(S(t)=\{1+(t/a)^p\}^{-1}\) Closed-form survival Decreasing or rises then falls AFT and proportional-odds interpretations

Here \(t>0\). Software uses incompatible names for scale, rate, and shape, so the implemented density or survival function must be checked rather than inferred from parameter labels.

The exponential distribution is nested in the Weibull family at \(p=1\). It is therefore a scientific hypothesis about a constant event rate, not merely a convenient default. Log-normal and log-logistic models allow a nonmonotone hazard, but neither can represent every biologically plausible shape.

3 Two Regression Interpretations

The same event-time distribution can be parameterized around different scientific comparisons. Coefficients must be interpreted from the chosen model, not from the word “survival.”

3.1 Proportional hazards

A proportional-hazards (PH) model has

\[ h(t\mid x)=h_0(t)\exp(x^\mathsf{T}\beta). \]

Holding other covariates fixed, a one-unit increase in \(x_j\) multiplies the instantaneous hazard by \(e^{\beta_j}\) at every time. A Weibull PH model supplies an explicit Weibull \(h_0(t)\); the Cox model leaves \(h_0\) unspecified. PH does not imply a constant hazard, only a constant hazard ratio.

3.2 Accelerated failure time

An accelerated-failure-time (AFT) model writes

\[ \log T=x^\mathsf{T}\gamma+\sigma\varepsilon, \]

where the distribution of \(\varepsilon\) determines the survival family. Its conditional quantiles obey

\[ \log Q_T(q\mid x+e_j)-\log Q_T(q\mid x)=\gamma_j, \]

or, equivalently, every modeled time quantile is multiplied by \(e^{\gamma_j}\) for a one-unit increase in \(x_j\). This time ratio is not a hazard ratio. A value \(e^{\gamma_j}=1.3\) means the modeled event-time scale is 30% longer, conditional on covariates.

Weibull models admit both PH and AFT parameterizations, with a sign and scale conversion between coefficients. Log-normal models are AFT but not PH. Software output should therefore state the parameterization and report the effect on the scale that answers the scientific question.

4 Likelihood with Incomplete Observation

Let \(Y_i=\min(T_i,C_i)\) and \(\delta_i=I(T_i\le C_i)\) for right censoring. Under conditionally independent censoring, subject \(i\) contributes

\[ L_i(\theta) =f(Y_i\mid x_i;\theta)^{\delta_i} S(Y_i\mid x_i;\theta)^{1-\delta_i}. \]

Thus an event contributes a density and a censored observation contributes the probability of surviving past its censoring time. With delayed entry at \(L_i\), the contribution is divided by \(S(L_i\mid x_i;\theta)\) because the observed sample is conditional on survival to entry. Interval censoring over \((a_i,b_i]\) contributes \(S(a_i)-S(b_i)\).

The maximum-likelihood estimator is the rule that maximizes the product of these contributions. Its observed value is an estimate of the selected distribution’s parameters, not a nonparametric fact about the population. Standard errors from the likelihood curvature require identification and regularity; sparse events, heavy censoring, boundary estimates, and flexible shapes can make Wald intervals unreliable. Profile likelihood or a design-respecting parametric bootstrap can better expose asymmetry, while still inheriting the fitted model.

5 Absolute Risk, Quantiles, and Extrapolation

Once \(S(t\mid x)\) is fitted, useful estimands include:

  • survival probability \(S(t^*\mid x)\) at a prespecified time \(t^*\);
  • event risk \(1-S(t^*\mid x)\);
  • median or another survival quantile \(Q_T(q\mid x)\); and
  • restricted mean survival time through \(\tau\), \(\operatorname{RMST}(\tau\mid x)=\int_0^\tau S(t\mid x)\,dt\).

Standardization over a target covariate distribution can turn conditional predictions into marginal population contrasts. This requires specifying the population, covariate distribution, intervention or exposure contrast, and time horizon.

Extrapolation beyond the last well-supported event is driven increasingly by the assumed tail. Several distributions can fit observed follow-up similarly yet predict very different long-term survival. Report the observed support, show numbers at risk, compare scientifically defensible families, and separate interpolation from extrapolation. External follow-up or mechanistic knowledge is more informative than selecting a tail solely by AIC.

6 Flexible Parametric Models

Simple named distributions restrict the hazard strongly. Flexible parametric survival models instead represent a transformation such as log cumulative hazard with restricted cubic splines. A common PH form is

\[ \log H(t\mid x)=s(\log t;\eta)+x^\mathsf{T}\beta, \]

where \(s\) is a spline. Interactions between covariates and functions of time can represent nonproportional hazards. Such models retain smooth absolute-risk prediction while relaxing a named baseline distribution.

Flexibility does not eliminate assumptions. Knot placement, boundary behavior, degrees of freedom, time-varying effects, and penalization affect the fit, especially in sparse tails. Model complexity must be judged against the number and timing of events rather than the number of rows or molecular features.

7 Diagnostics and Sensitivity

No single residual plot validates a parametric survival model. A useful workflow asks distinct questions.

  1. Audit time and observation. Check time zero, delayed entry, interval construction, competing events, censoring reasons, and covariate timing.
  2. Compare with nonparametric estimates. Overlay fitted survival curves with Kaplan–Meier estimates in meaningful groups, with uncertainty and risk-set sizes. Agreement in coarse groups does not validate covariate form.
  3. Check distributional shape. Use probability plots or cumulative-hazard transformations appropriate to the proposed family, and inspect Cox–Snell residuals with censoring retained.
  4. Check covariate functions and time variation. Examine nonlinear terms, interactions, residual patterns, and plausible time-varying effects.
  5. Assess influence and numerical stability. Inspect gradients, Hessian conditioning, profiles, starting values, and deletion sensitivity.
  6. Test conclusions across plausible models. Compare named and flexible distributions, censoring assumptions, time horizons, and tail constraints.
  7. Validate predictions. Evaluate time-specific calibration and prediction error with resampling or external data that repeats the complete modeling process.

Information criteria compare candidate likelihood models for the observed data; they do not establish correct extrapolation, causal interpretation, independent censoring, or transportability.

8 Bioinformatics Example: Time to Molecular Relapse

Suppose 320 patients contribute baseline tumor expression, clinical covariates, and time from surgery to molecular relapse. The patient is the experimental unit; assayed regions, cells, and reads are nested measurements. A log-normal AFT model might estimate how a prespecified expression score changes relapse-time quantiles, whereas a Weibull PH model targets a hazard ratio. Those are different estimands even when fitted to the same records.

The design should distinguish technical assay batches from biological patients, predefine how death before relapse is handled, prevent outcome-informed feature selection from leaking into validation, and retain center or treatment variables that affect observation and prognosis. If death prevents observing relapse, it is a competing risk, not ordinary independent censoring.

Statistically, a fitted time ratio describes the event-time distribution under the model and conditioning set. Biologically, it may reflect tumor aggressiveness, treatment allocation, surveillance intensity, comorbidity, or assay artifacts. Association alone does not identify a causal mechanism.

9 Connections and Take-home Messages

Parametric models convert explicit event-time assumptions into smooth absolute risk, quantile, and mean-survival estimates. Their coefficients may be hazard ratios or time ratios depending on parameterization. The reward is efficient estimation and possible extrapolation; the price is sensitivity to distributional shape, tail behavior, observation mechanisms, and correct biological replication.

10 References

Back to top