Survival Regression
Also known as: accelerated failure time model, AFT model, parametric survival model, time-to-event regression
Survival regression models the time until an event occurs — such as death, failure, or relapse — as a function of covariates. Unlike ordinary regression, it properly accounts for censored observations (cases where the event had not yet occurred at the end of follow-up) by specifying a parametric distribution for the survival time and estimating covariate effects via maximum likelihood.
Key highlights
- Handles right-censoring, left-censoring, and interval-censoring correctly, avoiding the bias of case deletion.
- Fully parametric specification enables extrapolation of survival curves and median survival time estimates beyond the observed study period.
- Accelerated failure time parameterisation provides directly interpretable time-ratio coefficients: a covariate multiplies the expected event time by a fixed factor.
- Likelihood-based estimation supports AIC/BIC model comparison for choosing among competing parametric families.
- Computationally efficient with closed-form gradients for common distributions.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use survival regression when the outcome is time to a well-defined event and censoring is present — clinical trials (time to relapse), engineering reliability studies (time to failure), and social science duration analysis (time to employment). It is preferred over Cox regression when a parametric distribution is plausible and interpretable acceleration factors (AFT coefficients) are desired, or when extrapolation beyond the observed follow-up is needed. Do not use it when events cannot be ordered in time, when the outcome is simply categorical or continuous without censoring, or when the assumed parametric family is a poor fit to the data (check with Cox-Snell plots).
Strengths & limitations
- Handles right-censoring, left-censoring, and interval-censoring correctly, avoiding the bias of case deletion.
- Fully parametric specification enables extrapolation of survival curves and median survival time estimates beyond the observed study period.
- Accelerated failure time parameterisation provides directly interpretable time-ratio coefficients: a covariate multiplies the expected event time by a fixed factor.
- Likelihood-based estimation supports AIC/BIC model comparison for choosing among competing parametric families.
- Computationally efficient with closed-form gradients for common distributions.
- Requires choosing a parametric distribution family; a mis-specified distribution can bias coefficient estimates and predictions.
- Extrapolation outside the observed time range is sensitive to the assumed tail shape and should be reported with caution.
- Less flexible than Cox regression for complex hazard shapes unless a mixture or piecewise model is used.
- Large samples are needed when the event rate is low, since each parameter is estimated from relatively few events.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
How is survival regression different from Cox regression?
Cox regression is semi-parametric: it models covariate effects without specifying the baseline hazard shape, which makes it robust but prevents extrapolation. Survival regression is fully parametric: it specifies the hazard family (e.g., Weibull), enabling extrapolation and yielding interpretable AFT time-ratio coefficients at the cost of requiring a distributional assumption.
What is censoring and why does it matter?
Censoring occurs when the event time is only partially observed — a patient leaves the study before experiencing the event, so only a minimum event time is known. Dropping censored observations biases results toward shorter survival times. Survival regression handles censoring correctly by including censored cases in the likelihood through the survival function S(t).
How do I choose the right parametric distribution?
Compare competing distributions using AIC or BIC, and check fit with Cox-Snell residual plots: well-fitted models produce residuals that follow an exponential(1) distribution. The Weibull family is a common default because it nests the exponential and allows monotone hazard shapes; log-normal or log-logistic suit humped or declining hazard patterns.
Can survival regression handle time-varying covariates?
Standard parametric survival regression assumes fixed covariates. Time-varying covariates can be accommodated by reformatting the data into interval-split episodes and using a piecewise exponential or similar model, but this substantially increases data complexity.
What is an accelerated failure time (AFT) model?
An AFT model re-parameterises the survival regression so that covariates multiply (or divide) the expected survival time rather than the hazard. A time-ratio coefficient of 1.5, for instance, means the covariate extends median survival by 50%. This is often more interpretable than a hazard ratio in clinical or policy contexts.
Sources
- 1.Kalbfleisch, J. D., & Prentice, R. L. (2002). The Statistical Analysis of Failure Time Data (2nd ed.). Wiley.ISBN 978-0471363576
- 2.Cox, D. R., & Oakes, D. (1984). Analysis of Survival Data. Chapman and Hall.ISBN 978-0412244902
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Survival Regression. ScholarGate. https://scholargate.app/statistics/survival-regression