Accelerated Failure Time Model

Accelerated failure time model (AFT model) is a major class of parametric regression models in biostatistics, reliability engineering, and infectious-disease epidemiology that relates explanatory covariates directly to the logarithm of survival time T rather than to the instantaneous hazard rate. Whereas the semi-parametric proportional hazards model assumes that covariates multiply a baseline hazard function by a time-invariant hazard ratio, the accelerated failure time framework posits that physiological stressors, genetic risk variants, or therapeutic interventions linearly accelerate or decelerate the speed at which an organism progresses along its biological failure trajectory. Developed across industrial life-testing and biometric theory by David R. Cox (1972), Ross L. Prentice (1978), and John D. Kalbfleisch (1980), AFT models yield regression coefficients that translate directly into time ratios (or acceleration factors), providing clinicians and public health planners with an intuitive physical measure of how much a treatment expands median or quantile survival time.

Multiplicative Time Scaling and Survival Function Geometry

Let T_i > 0 denote the time from a defined origin to the event of interest for individual i (i = 1, …, n) with a p-dimensional covariate vector x_i = (x_i1, …, x_ip)^⊤. On the natural logarithmic scale, the accelerated failure time model specifies a linear location-scale regression:

where β₀ is the baseline intercept, β = (β₁, …, β_p)^⊤ is the vector of covariate log-time ratios, σ > 0 is a positive scale parameter governing dispersion, and ε_i is an independent random error term following a specified baseline probability distribution with survival function S_ε(ε). Exponentiating both sides expresses failure time T_i as a multiplicative perturbation of a baseline random lifetime T_0i = exp(β₀ + σ ε_i):

Let φ = exp(−x_i^⊤ β) denote the acceleration factor (and its reciprocal ψ = exp(x_i^⊤ β) denote the time ratio, TR). Under this parametrization, the survival function S(t | x_i) = Pr(T_i > t | x_i) and instantaneous hazard function h(t | x_i) are related to the baseline survival function S₀(t) and baseline hazard h₀(t) (where x = 0) by:

Physical Interpretation of Time Ratios and Quantile Invariance

Because the covariate factor exp(x_i^⊤ β) scales time t directly inside S₀(·), any survival quantile t_q(x) (such as the median survival time t_0.50 where S(t_0.50) = 0.50, or the 25th percentile assessed on a Kaplan–Meier estimator) shifts by the same constant multiplicative ratio across covariate profiles:

• If β₁ > 0 (so TR = exp(β₁) > 1 and acceleration factor φ < 1), the exposure decelerates failure progression and prolongs survival. For instance, exp(β₁) = 1.60 implies that median survival—and every other survival percentile—is 60% longer in the treated group than in the control group.
• If β₁ < 0 (so TR = exp(β₁) < 1 and φ > 1), the exposure accelerates biological aging or disease progression, compressing time-to-event by factor TR (analogous to running an physiological clock at φ times normal speed).

By choosing the distribution of the standardized error term ε_i = [ln(T_i) − (β₀ + x_i^⊤ β)] / σ, the AFT framework generates a rich hierarchy of parametric survival distributions capable of modeling monotonic, unimodal (humped), or bathtub-shaped hazard trajectories.

The Weibull Dual-Representation Property

The Weibull distribution occupies a unique mathematical position in survival analysis: it is the only continuous parametric family (including its exponential special case σ = 1) that is simultaneously closed under both the accelerated failure time assumption and the proportional hazards assumption. Specifically, a Weibull AFT coefficient β_AFT with scale parameter σ corresponds to a Cox proportional hazards log-hazard ratio β_PH via the identity:

Log-Logistic Proportional Odds and Generalized Gamma Nesting

When the error term ε_i follows a standard logistic distribution with cumulative distribution function F_ε(u) = 1 / (1 + e^−u), the resulting log-logistic AFT model satisfies the proportional odds property for survival: the odds of surviving beyond time t, defined as S(t | x) / [1 − S(t | x)], are multiplied by exp(x^⊤ β / σ) across all t > 0. Consequently, the hazard ratio between two groups converges toward 1.0 as t → ∞, making the log-logistic AFT model well suited for diseases in which initial susceptibility differences attenuate among long-term survivors.

Because the three-parameter generalized gamma distribution introduced by E. W. Stacy (1962) and parameterized for censored regression by Ross L. Prentice (1974) nests the Weibull (q = 1), log-normal (q = 0), and standard gamma (q = σ) distributions as special cases, biostatisticians use nested likelihood-ratio tests within the generalized gamma AFT model to select the appropriate parametric family empirically rather than imposing an arbitrary hazard shape.

Censored Maximum Likelihood and Buckley–James Semiparametric Estimators

For a sample of n subjects with observed follow-up times y_i = min(T_i, C_i) and event indicators δ_i = I(T_i ≤ C_i) under independent right-censoring (or interval censoring [L_i, R_i]), parametric AFT parameters (β, σ) are estimated by maximizing the censored log-likelihood:

where z_i = [ln(y_i) − β₀ − x_i^⊤ β] / σ, and f_ε(·) and S_ε(·) are the density and survival functions of ε. When investigators wish to avoid specifying f_ε(·) parametrically, semi-parametric AFT estimation can be performed using either the Buckley–James estimator (1979), which iteratively imputes right-censored log-survival times via the Kaplan–Meier estimator of the residuals, or rank-based linear regression estimators (Tsiatis 1990; Wei, Ying, and Lin 1990) derived from weighted log-rank test estimating equations.

Robustness to Unobserved Heterogeneity (Omitting Covariates)

A major methodological advantage of the AFT model over the Cox proportional hazards model concerns collapsibility and omitted-covariate bias. As demonstrated in frailty models, omitting a balanced prognostic covariate (or unmeasured individual susceptibility Z) from a proportional hazards model causes the marginal cohort hazard to decelerate over time, violating proportional hazards and biasing the estimated hazard ratio toward the null (1.0) even in a randomized trial. By contrast, in the linear log-time formulation ln(T_i) = β₀ + x_i^⊤ β + γ U_i + σ ε_i, an unmeasured covariate U_i independent of x_i is absorbed directly into the composite error term γ U_i + σ ε_i without biasing the point estimate of the treatment time ratio exp(β).

1. Infectious-Disease Incubation Periods: Since Philip E. Sartwell (1950) demonstrated that the incubation periods of acute infectious diseases closely follow log-normal distributions, parametric log-normal, gamma, and Weibull AFT models with interval censoring have served as the standard statistical architecture for estimating pathogen incubation quantiles (median, 95th, and 97.5th percentiles used to set quarantine durations) and evaluating whether host age or viral variant accelerates symptom onset.
2. Oncology Trials with Non-Proportional Hazards and Extrapolation: In health technology assessment and cost-effectiveness modeling (where lifetime survival beyond trial follow-up must be integrated alongside restricted mean survival time or Royston–Parmar flexible parametric models), AFT models provide closed-form tail projections and direct time-expansion metrics across patient subgroups.

1. Cox, D. R. (1972). "Regression models and life-tables." Journal of the Royal Statistical Society: Series B (Methodological), 34(2), 187–220. https://doi.org/10.1111/j.2517-6161.1972.tb00899.x
2. Prentice, R. L. (1974). "A log gamma model for survival data." Biometrika, 61(3), 539–544. https://doi.org/10.1093/biomet/61.3.539
3. Buckley, J., & James, I. (1979). "Linear regression with censored data." Biometrika, 66(3), 429–436. https://doi.org/10.1093/biomet/66.3.429
4. Wei, L. J. (1992). "The accelerated failure time model: A useful alternative to the Cox regression model in survival analysis." Statistics in Medicine, 11(14–15), 1871–1879. https://doi.org/10.1002/sim.4780111409
5. Kalbfleisch, J. D., & Prentice, R. L. (2002). The Statistical Analysis of Failure Time Data (2nd ed.). John Wiley & Sons. https://doi.org/10.1002/9781118032985