Does an AFT model require every event time to be observed?

No. Suitable AFT estimation methods can use censored observations, subject to assumptions about the censoring process.

Does a time ratio of 1.5 mean a 50% lower failure risk?

No. It means a 50% increase in corresponding event-time quantiles under the stated AFT model. Risk at a particular time is a different quantity.

Must every AFT model be Weibull?

No. Weibull is one option; lognormal, log-logistic and other formulations are also used. The distribution choice affects the model's shape and predictions.

AI Glossary / Statistics and probability / Intermediate

Accelerated failure time model

Definition

An accelerated failure time model, or AFT model, relates explanatory variables to the time until an event by stretching or compressing the time scale. In a common formulation, the logarithm of event time is modeled as a linear function of covariates plus an error term. Exponentiated coefficients describe time ratios under the stated parameterization.

What it means

The event need not be a physical failure. It could be cancellation of a subscription, completion of a process or another clearly defined occurrence. The name comes from survival and reliability analysis, where researchers often study time to failure.

An AFT model asks how event-time distributions change with explanatory variables. A time ratio above one means longer times under the conventional log-time formulation used here; a ratio below one means shorter times. Whether a longer time is desirable depends on the event. Delaying equipment failure and delaying service completion have different business meanings.

How it works

A basic formulation is log(T) = beta0 + beta1*x1 + ... + sigma*error, where T is positive event time. Holding other variables fixed, increasing x1 by one unit multiplies corresponding event-time quantiles by exp(beta1), provided the error distribution and scale remain unchanged.

Common parametric choices include Weibull, lognormal and log-logistic event-time distributions. AFT methods also have semiparametric forms. The model must handle censoring appropriately: if equipment is still operating when observation ends, its recorded duration is a lower bound on failure time, not an observed failure.

Fit and interpretation depend on the event definition, covariate timing, censoring assumptions and distributional adequacy. A time ratio is not generally a hazard ratio. Some special distributions support both AFT and proportional-hazards representations, but the two model families should not be treated as interchangeable.

Three practical examples

1. A reliability analyst estimates how operating conditions are associated with equipment lifetime while retaining units that have not yet failed as censored observations.

2. A subscription service models time to cancellation to support planning, distinguishing observed cancellations from accounts still active when the data window closes.

3. An operations team studies time to complete a workflow, using a clearly defined start and completion event rather than treating unfinished cases as completed at the reporting date.

Think of it this way

Think of two versions of the same timeline played at different speeds. Stretching the timeline moves corresponding events later; compressing it moves them earlier. The AFT assumption says that covariates act through this kind of time rescaling, rather than imposing an arbitrary change to every part of the distribution.

Worked example

Consider an original illustrative AFT model for component lifetime. Let x be zero for a reference operating condition and one for another condition, with log(T) = log(T0) + log(1.5)*x. The time multiplier for x = 1 is therefore 1.5.

Suppose the reference median lifetime is 400 hours and its 25th event-time percentile is 240 hours. Under the shared-shape AFT assumption, the corresponding values for x = 1 are 1.5 x 400 = 600 hours and 1.5 x 240 = 360 hours. The whole distribution is rescaled, not just its median.

A component still running at 500 hours is right-censored at that point. Its failure time is known to exceed 500 hours; entering 500 as an observed failure would misrepresent the data. These numbers are assumed for explanation and do not demonstrate a causal effect of changing operating conditions.

Illustrative case study

Fictional case study: Timewell Services studies the duration of maintenance jobs. Several jobs remain open at the end of the reporting period, and an initial spreadsheet records the reporting date as if it were each job's completion date.

The analyst corrects the event indicator and treats those jobs as right-censored. The team then checks whether a common time-rescaling assumption is plausible across job types, rather than interpreting every fitted coefficient immediately. It also reviews whether difficult jobs are disproportionately missing from follow-up.

This fictional case illustrates why event recording and censoring matter before model interpretation. It does not claim a successful fit or a reduction in completion time, and a fitted association would not alone establish how a process change would affect future jobs.

Common misconceptions and mistakes

The word "accelerated" does not mean the model always predicts faster events. Positive coefficients in the stated log-time convention extend time. Other software or engineering definitions may use reciprocal acceleration factors, so check the formula before interpreting a sign or ratio.

Do not confuse a time ratio with a risk difference or hazard ratio. Also, excluding every unfinished observation can bias the analysis. AFT modeling accommodates censoring through a suitable estimation method; it does not automatically solve informative loss to follow-up or an incorrect event definition.

Business use

AFT models can make duration analyses easier to explain because multiplicative changes in time are often directly relevant to planning. Potential applications include reliability, customer retention and process analysis, provided the outcome and observation scheme fit the model.

Report the time-ratio convention, uncertainty intervals and the population to which the estimate applies. Check model fit and predictions on relevant time ranges. Extrapolating far beyond observed durations is particularly sensitive to the chosen distribution. For decisions about interventions, separate predictive associations from causal conclusions and examine how the data were collected.

Check your understanding

Under log(T) = log(T0) + log(0.8)*x, the reference median is 100 days. What median is implied when x changes from zero to one?

Answer: The implied median is 0.8 x 100 = 80 days. This is a shorter event time, not automatically a beneficial outcome.

Related terms

Survival analysis
Censoring
Proportional hazards
Weibull distribution
time ratio
lognormal distribution
survival function and accelerated life testing
assumptions and experimental settings

Learn next

compare survival curves related by a constant time multiplier, then examine curves whose shapes differ. Check how an event indicator changes the contribution of a completed versus censored observation before interpreting regression coefficients.

Regression coefficients

Sources

Primary sources: lifelines, Survival regression, accelerated failure time models: https://lifelines.readthedocs.io/en/latest/Survival%20Regression.html

lifelines, WeibullAFTFitter model definition: https://lifelines.readthedocs.io/en/latest/fitters/regression/WeibullAFTFitter.html

NIST/SEMATECH, Common acceleration models and acceleration-factor conventions: https://www.itl.nist.gov/div898/handbook/apr/section1/apr15.htm

Previous: ABX test Next: Accented syllables