Description
Survival model in which covariates speed up or slow down time to an event by a multiplicative factor, with a chosen baseline distribution and censored cases handled in the likelihood.
Durations (time between purchases, time to cancel, time to respond) are positive and often censored, so they need hazard-based models. In the proportional hazards model covariates multiply the hazard; in the accelerated failure time model they stretch or shrink time itself through a factor of exp(-X beta), so coefficients read as percentage changes in expected duration, which stakeholders often find more intuitive. The two coincide only for the Weibull distribution. Baselines include exponential (memoryless), Weibull (monotone rising or falling hazard), and loglogistic or lognormal (hazards that rise then fall). Censored spells contribute their survival probability to the likelihood, and time-varying covariates can be included if they change a limited number of times. Fit is checked with generalised residuals, which should be unit exponential: plot their empirical integrated hazard against a 45-degree line and run conditional moment tests. Unobserved heterogeneity can be added but is hard to separate from duration dependence.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Siloed
- Stage
- Post-launch
- Data
- Quantitative
- Effort
- Days
Steps
- Define the spell (start and end), the event and a censoring indicator for each case.
- Choose a baseline distribution (Weibull, loglogistic, lognormal) and the accelerated failure time or proportional hazards form.
- Estimate by maximum likelihood with density terms for completed spells and survival terms for censored ones.
- Compute generalised residuals, plot their empirical integrated hazard against the 45-degree line and run moment tests; move to a more flexible baseline if rejected.
- Compare models with likelihood ratio tests and AIC or BIC.
- Interpret coefficients as percentage changes in expected duration and forecast durations or the number of events per period on a hold-out sample.
Sources
- Quantitative Models in Marketing Research (Franses & Paap, 2001)
- Applied Bayesian Modelling (Congdon, 2003)
- Statistical Methods for Reliability Data (Meeker & Escobar, 1998)
- Statistical Methods for Survival Data Analysis (Lee & Wang, 2003)
- Bayesian Statistical Modelling (Congdon, 2006)
Also known as
- Accelerated Lifetime model
- AFT model
- Parametric survival regression (Weibull, loglogistic)
- Accelerated lifetime (accelerated failure time) duration model
Categories
- Analytics & Behaviour
- Statistics & Modelling