The accelerated failure time model helps you predict when an event, such as equipment failure or customer churn, will occur. Unlike methods that only look at relative risk, this approach estimates how specific factors speed up or slow down the time until an outcome happens, providing a clear timeline.
Understanding the Time Ratio
The core output of the accelerated failure time model is the time ratio. If a factor has a time ratio greater than one, it implies the factor stretches out the expected time until an event, essentially acting as a 'slow-down' effect. Conversely, a ratio less than one indicates that the factor accelerates the process, leading to a faster occurrence of the event.
This measurement is highly valuable for planning, such as estimating how long a machine will last under different operating conditions or how long a customer will remain active based on their subscription tier.
Handling Parametric Assumptions
Unlike non-parametric or semi-parametric models, this method assumes that your survival data follows a specific mathematical distribution. By fitting a parametric model, you can make more precise predictions about the future, even beyond the timeframe of your current data.
Lattice automatically performs a post-check to ensure that the chosen distribution fits your data well. If the model struggles to converge or the shape parameters fall into unrealistic ranges, the platform will alert you to ensure your conclusions remain reliable.
Interpreting Model Convergence
A successful analysis depends on the model converging correctly. If the model encounters issues—such as an inability to find a stable solution for your specific dataset—the platform will flag this as a 'convergence concern'.
This transparency prevents misleading results. When you see this warning, it suggests that the underlying data might be too sparse or noisy for a full parametric model, and you may need to reconsider your grouping or check for data quality issues before proceeding.
1 · Intent → method
An LLM picks survival_aft from a fixed catalog.
2 · Method → numbers
Deterministic Python engine runs the math. Same input → same output.
3 · Numbers → plain language
A second LLM translates the result into your domain’s vocabulary.
How is this different from a standard survival regression like Cox PH?
While Cox regression focuses on the relative hazard ratio, the accelerated failure time model estimates the 'time ratio'. This allows you to say, for example, that a treatment increases the expected survival time by a specific factor, providing a more intuitive sense of duration.
What do the different distributions like Weibull or Log-Normal represent?
These distributions represent different assumptions about how failure rates change over time. The accelerated failure time model on Lattice allows you to choose these to best fit your data's specific aging or survival pattern, such as whether failure rates increase or decrease as time progresses.
Tool input schema
Schema for survival_aft not exported yet (run pnpm export:registry).