vldaft

Varying Location and Dispersion Accelerated Failure Time (VLDAFT) regression, implementing the model from Anderson (1991). Both the location (μ) and the dispersion (σ) of the log-survival time can depend on covariates, and the dispersion can depend on the fitted location through a polynomial coupling — enabling non-proportional hazards under the Weibull and other parametric distributions. The C code that supported the original 1991 manuscript was updated to the standards required for an R package. Either a C or a Rust back end can be used for the computations.

Installation

# install.packages("pak")
pak::pak("keaven/vldaft")

System requirements

A Rust toolchain (rustc + cargo) is required. Install from https://rustup.rs.

Usage

The Framingham Heart Study data used in Anderson (1991) are confidential and are not shipped with this package. They live in the companion private repository keaven/vldaft.data, which also carries the regression tests that depend on them.

The example below uses the freely available lung dataset from the survival package so that anyone can run it out of the box.

library(vldaft)
library(survival)

# Constant-dispersion (standard Weibull AFT)
fit1 <- vldaft(
  Surv(time, status) ~ age + sex + ph.ecog,
  data  = na.omit(lung),
  dist  = "weibull",
  theta = 0
)

# Dispersion coupled to location via a linear theta polynomial
fit2 <- vldaft(
  Surv(time, status) ~ age + sex + ph.ecog,
  data  = na.omit(lung),
  dist  = "weibull",
  theta = 1
)

summary(fit2)

# Likelihood-ratio test for "is dispersion proportional to location?"
2 * (fit2$loglik - fit1$loglik)

Features

• Five error distributions: Weibull, logistic, normal, Cauchy, gamma.
• Censoring: right censoring, left censoring, left truncation.
• Scale–location coupling: polynomial theta parameters linking log(sigma) to mu.
• Scale covariates: separate covariates for the scale (dispersion) model via formula syntax Surv(...) ~ loc_vars | scale_vars.
• Dual back ends: C and Rust implementations, selectable via backend = c("c", "rust").
• gamlss2 families: AFT_Weibull(), AFT_Logistic(), AFT_Normal(), AFT_Cauchy(), AFT_Gamma().

Reference

Anderson, K. M. (1991). A nonproportional hazards Weibull accelerated failure time regression model. Biometrics, 47(1), 281–288.