Statistics Glossary → Parametric Survival (AFT)
Parametric Survival (AFT)analysis
A survival model that describes the time until an event with a specific probability distribution, such as Weibull or log-normal, and gives by what factor each factor lengthens or shortens that time (the time ratio).
Concept
Cox regression says by what factor a variable changes the risk of the event and never specifies how the risk varies over time. Parametric survival models instead assume that the time to the event follows a specific distribution. If this assumption is close to the truth, the model estimates more efficiently, can make reasonable predictions beyond the follow-up period, and gives results directly on the time scale: 'this factor roughly halves survival time'.
In the accelerated failure time (AFT) model, the logarithm of time is a linear function of the covariates: log T = β₀ + β₁x₁ + … + σW. The distribution of W sets the type of model (Weibull, log-normal, log-logistic, exponential, generalised gamma). The exponential of a coefficient, exp(β), is the time ratio (TR): TR = 2 says that a one-unit increase in that factor doubles the time to the event; TR = 0.5 halves it. The factor in effect changes how fast time runs. The shape of the distribution sets whether the hazard stays constant over time (exponential), rises or falls steadily (Weibull) or first rises and then falls (log-normal, log-logistic).
The results depend on the choice of distribution, so different distributions should be compared with AIC and BIC and the model curve checked visually against the Kaplan-Meier curve. The Weibull and exponential distributions can be read both as AFT and as proportional hazards models; for the Weibull, the time ratio can be converted into a hazard ratio. The assumption of non-informative censoring applies here as it does in Kaplan-Meier and Cox.
When to use it
- When modelling time to an event with censored data and you want to express results on the time scale (by what factor time is lengthened or shortened), estimate percentile times or predict beyond the follow-up period.
- In engineering and reliability studies, when product lifetimes are thought to follow a known distribution (most often Weibull).
- If you do not want to assume anything about the shape of the hazard over time, use Cox regression; if the event time is known only to lie within an interval, interval-censored survival models are preferred.
An everyday example
A battery maker follows the time to failure of 400 car batteries sold in two climate zones; when the study ends some batteries are still working (censored). If a Weibull AFT model gives a time ratio of 0.75 for the hot zone, a battery's life there is about 75% of that in the cool zone. The model might also estimate a median life of 5 years in the cool zone and 3.75 years in the hot zone, and the maker can set the warranty period accordingly.
Parameters in MerQur
Time column, Event (0/1) and Covariates
Concept: Time is the period from the start of follow-up to the event or to censoring; the event indicator says whether this time ended with the event (1) or with censoring (0). Covariates are the variables thought to affect the time to the event.
Everyday example: In the battery example, time is the number of months from sale to failure or to the end of the study, the event is 'did it fail?', and the covariates are climate zone and the vehicle's daily usage time.
What it does: The response and covariates of the model log T = xβ + σW; text covariates are coded with K-1 dummy variables. The covariate selector also lists text categoricals. At least 30 observations and 10 events are required.
Effect on the result: The whole model changes.
On the validation data: Coefficients and likelihoods are given under the distributions below.
Event codes (optional)
Concept: If the status column is coded other than 0/1, states which values count as events; the other values are treated as censored.
Everyday example: If the service records say '1 = working' and '2 = failed', the event code 2 is entered; batteries coded 'working' are counted as censored.
What it does: Specifies the values counted as events; the others are censored.
Effect on the result: Rebuilds the event indicator.
On the validation data: lung "2": ph.ecog -0.3557, sex 0.4014, Intercept 5.8196, σ = 0.7310. When empty, lung 1/2 raises an error. In the headache data, if the event code is left empty, Censor = 1 is counted as the event; with 8 events the 'at least 10 events' rule stops the model. With event code 0 the model is fitted with 30 events.
Categorical (factor) variables
Concept: Lets a variable that is coded with numbers but is really categorical (e.g. regions coded 1, 2, 3) be modelled with dummy variables instead of a single numeric slope; each level is read relative to the base level.
Everyday example: If climate zone is coded 1 = cool, 2 = mild, 3 = hot and is not made a factor, the model assumes, meaninglessly, that 'each step up in zone number changes life by a fixed amount'. As a factor, the mild and hot zones are each compared with the cool zone separately.
What it does: Makes a numerically coded column a factor (K-1 dummies; baseline level alphabetical or numerically first level, or the level entered in the Reference levels field).
Effect on the result: The number of coefficients and their interpretation change.
On the validation data: larynx "stage" as a factor: stage_2 -0.1477, stage_3 -0.5866, stage_4 -1.5441. The same model with baseline level "stage=4": stage_1 1.5441, stage_2 1.3964, stage_3 0.9575.
Distribution (Weibull / Log-Normal / Log-Logistic / Exponential / Generalized Gamma / Gompertz (PH))
Concept: Sets which distribution time follows and, with it, how the hazard changes over time: constant for the exponential, steadily rising or falling for the Weibull, possibly first rising then falling for the log-normal and log-logistic, rising exponentially for the Gompertz; the generalised gamma is a flexible family that contains several of these as special cases.
Everyday example: If batteries become more likely to fail as they age, the Weibull may suit. If some new batteries fail in the first months because of manufacturing defects and the risk then falls, the log-normal or log-logistic may fit better.
What it does: Selects the error distribution. Weibull, log-normal and log-logistic run with the lifelines AFT engine; exponential, generalized gamma and Gompertz run with MerQur's own maximum likelihood estimation (analytic Hessian in exponential, Richardson numerical Hessian in Gompertz and generalized gamma).
Effect on the result: All coefficients, scale and likelihood change. Weibull "rho" = 1/σ (the Weibull shape parameter), log-logistic "beta" = 1/σ, and log-normal is reported as σ.
On the validation data: larynx, z1 z2 z3 age: LL Weibull -141.4234, log-normal -141.3917, log-logistic -141.5890, exponential -141.8978, gen. gamma -141.0769, Gompertz -141.3783; time ratio for z3 0.2135 / 0.1561 / 0.1710 / 0.1949 / 0.1774, Gompertz HR 5.8261.
→ Accelerated failure time (AFT) model and survival distributions
Reference levels
Concept: For a categorical variable coded as text, sets which level is the comparison standard (base level). Changing the base level does not change the model's fit; only which group the coefficients are read against.
Everyday example: If zone is coded with the texts 'cool, mild, hot' and you want results read against the hot zone, 'hot' is made the base level. The cool zone's time ratio then comes out above 1: 'batteries live longer in the cool zone'.
What it does: Sets the baseline (comparison) level of a text covariate.
Effect on the result: The dummy coefficients and the intercept change.
On the validation data: headache (with event code = 0) with 'Group=grup2': Group_grup1 -0.1933, Intercept 3.3091 (baseline level = last level); when empty, grup1 is the baseline level (Group_grup2 +0.1933, SE 0.0786, Intercept 3.1158; rho 4.7128, AIC 205.75). mgus2 with "sex=M": sex_F +0.3187.
Confidence Level (%90 / %95 / %99 / Other (custom))
Concept: Sets the confidence level of the interval estimates for the coefficients and time ratios. If a time ratio's interval excludes 1, the factor changes time significantly.
Everyday example: If the hot zone's time ratio is 0.75 with a 95% confidence interval of 0.65-0.87, batteries clearly live shorter there; choosing 99% widens the interval, which may still exclude 1.
What it does: The level of the coefficient and time ratio intervals. The significance mark is given with the selected α.
Effect on the result: Only the intervals change.
On the validation data: larynx z3 time ratio interval: 95% (0.1048, 0.4351), 90% (0.1175, 0.3881), 99% (0.0838, 0.5442), 80% (0.1340, 0.3401), 99.9% (0.0646, 0.7056).
Additional output: AIC/BIC table across distributions
Concept: Fits each of the six distributions with the same covariates and shows their log-likelihood, AIC and BIC side by side. The smallest AIC points to the distribution with the best balance of fit and complexity; if the AIC differences are small, the distributions fit similarly in practice.
Everyday example: If the AICs of the Weibull, log-normal and log-logistic models are within 1-2 units of each other, the data do not force a clear choice among them; if their time ratios are also similar, the conclusion is robust to the choice of distribution.
What it does: Gives the number of parameters, LL, AIC, BIC and ΔAIC for the six distributions (the smallest AIC is marked).
Effect on the result: Only adds output.
On the validation data: larynx: exponential has the smallest AIC (293.80; ΔAIC Weibull 1.05, gen. gamma 2.36). whas100: the gen. gamma row comes out with a degenerate fit LL = -102.88; the boundary solution is marked with "(!)" and excluded from the best choice (best: Weibull).
Additional output: Weibull PH parametrization (HR, σ)
Concept: The Weibull model is at once an AFT and a proportional hazards model. This output converts time ratios into hazard ratios like those of Cox; a shape parameter above 1 means the hazard rises over time, below 1 that it falls.
Everyday example: With a time ratio of 0.75 for the hot zone, the hazard ratio is about 1.5 (with a shape parameter of 1.5): the same finding is stated as 'batteries live 25% shorter' or 'the risk of failure at any moment is about 1.5 times higher'. A shape of 1.5 means the failure risk rises as batteries age.
What it does: For Weibull only, log HR = -b/σ (SE by the delta method), the σ interval, rho and lambda.
Effect on the result: Only adds output; produces no output when another distribution is selected.
On the validation data: larynx z3 log HR = 1.745 (HR 5.726), σ = 0.8848, rho = 1.1301, lambda = 0.018537.
→ Accelerated failure time (AFT) model and survival distributions
Additional output: Predicted percentiles and S(t) (at covariate means)
Concept: The model predicts, for a unit at average covariate values, percentiles of time (e.g. when 10% or 50% of units have had the event) and survival probabilities at chosen times. Because a parametric distribution is used, these predictions can be extended beyond the follow-up period, but must be interpreted with care there.
Everyday example: The maker may ask 'after how many months have 10% of batteries failed?'; if the answer is 22 months, the warranty can be kept below that. The median life is the time by which half the batteries have failed.
What it does: Gives the 10th, 25th, 50th, 75th and 90th percentiles and S(t | x mean) values at the covariate means.
Effect on the result: Only adds output.
On the validation data: motorette (single covariate) 50% = 1614.96; larynx Weibull 50% = 5.2028; percentiles are also given for the exponential, gen. gamma and Gompertz distributions and for the model with a factor.
Example on real data
Data: Larynx cancer, Klein and Moeschberger 2003, R KMsurv
Variables: time time (years), event delta (death), covariates z1, z2, z3 (stage dummies) and age
- Question: In patients with laryngeal cancer (n = 90, 50 deaths), how much do disease stage (z1, z2, z3: dummies for stages II, III, IV; baseline level stage I) and age shorten survival time? (Weibull AFT model)
- Model test: LR χ²(4) = 19.37, p = 0.00066; the covariates are jointly significant. Log-likelihood -141.42, AIC 294.85.
- Result: Only stage IV is significant: the z3 coefficient is -1.5441 (SE 0.3633, p < .001), time ratio TR = 0.2135 [0.105, 0.435]; the survival time of stage IV patients is shortened to about one fifth of stage I. Stage III is borderline (TR = 0.556, p = 0.067); stage II and age are not significant.
- Distribution parameter: Weibull shape rho = 1.1301 (> 1); the hazard increases slightly over time.
- Fit: The model survival is close to the KM curve (0.6632 and 0.7108 at year 3; 0.3082 and 0.2965 at year 9); the choice of distribution can be compared with the log-normal or log-logistic model by AIC.
- APA: In the Weibull AFT model, stage IV significantly shortened survival time, b = -1.54, SE = 0.36, TR = 0.21, 95% CI [0.11, 0.44], p < .001; LR χ²(4) = 19.37, p < .001.
Related concepts
Percentiles (quantiles) and percentile definitionsAccelerated failure time (AFT) model and survival distributionsReference category (comparison level)Information criteria (AIC, AICc, BIC)Dummy (indicator) variable codingCovariateError distribution (normal, Student-t)
The numbers in the examples are MerQur's output on independent, publicly available validation data; see the MerQur Technical Note for details.