Statistics Glossary → Accelerated failure time (AFT) model and survival distributions
Accelerated failure time (AFT) model and survival distributionsconcept
A parametric survival model in which covariates stretch or shrink the time to the event by a multiplicative factor, with log(time) described by a linear model and a chosen error distribution (Weibull, log-normal, log-logistic and others).
The AFT model has the form log(T) = β₀ + β₁x₁ + ... + σ·ε, where the distribution of ε determines the distribution of the times: an extreme-value distribution gives Weibull times (with the exponential as a special case), a normal distribution gives log-normal times and a logistic distribution log-logistic times. exp(β) is a time ratio (acceleration factor): exp(β) = 1.5 means that a one-unit increase in the covariate multiplies the time to the event by 1.5, that is, delays the event. This interpretation can be more intuitive than the hazard ratio of the Cox model.
The Weibull is the only family that can be written both as an AFT and as a proportional hazards model (the exponential being its special case), so Weibull coefficients can also be converted into hazard ratios (log HR = -β / σ). The log-normal and log-logistic distributions allow hazards that first rise and then fall, and do not imply proportional hazards. The Gompertz distribution is usually used in proportional hazards form. When the right distribution is chosen, parametric models are more efficient than the Cox model and allow prediction beyond the follow-up period, but the wrong distribution gives biased results; distributions should be compared with AIC and with plots.
An everyday example
A light-bulb maker tests two filament materials. In an AFT model the coefficient for the new material gives exp(β) = 1.3: bulbs with the new material burn 30% longer on average. If the median life with the old material is 1,000 hours, with the new one it is about 1,300 hours; it is as if time runs more slowly for the new material.
In MerQur
It appears as the 'Distribution' option (Weibull, Log-Normal, Log-Logistic, Exponential, Generalized Gamma, Gompertz) and 'Additional output: Weibull PH parametrization (HR, σ)' in Parametric Survival, and as 'Regression model' (Weibull, Log-normal, Log-logistic or semi-parametric PH) in Interval-Censored Survival.
Analyses where this concept appears
Parametric Survival (AFT)Interval-Censored Survival