`flexsurvspline`: optimization errors when using left-truncation for age at study entry

[jarbet]

Related to #110

I'm trying to use left-truncation to account for age at study entry, and use age as the underlying time-scale, as recommended here).

suppressPackageStartupMessages({
  library(survival)
  library(flexsurv)
})
packageVersion("flexsurv")
#> [1] '2.3.2'

pbc$start <- pbc$age
pbc$stop <- pbc$age + pbc$time / 365.25

fit <- flexsurvspline(
  Surv(start, stop, status == 2) ~ sex,
  data = pbc,
  k = 4
)
#> Error in optim(method = "BFGS", par = c(gamma0 = 0, gamma1 = 0, gamma2 = 0, : initial value in 'vmmin' is not finite

^{Created on 2026-03-09 with reprex v2.1.1}

[chjackson]

This is a failing of the procedure that gets initial values for optimisation in spline models, and needs more
investigation.

You may be able to work around it in the meantime by supplying initial values for the parameters. All that's needed is to find parameters that give a non-zero likelihood. To find these, you can often start with a simpler model that works, then use the estimates as initial values for the same parameters in the more complex model. Here I came up with the -12, 3 for gamma0 and gamma1 by fitting a model with k=0, which works here. Then initialising all the other parameters to zero gives a valid starting value for optimisation with k=4.

fit <- flexsurvspline(
  Surv(start, stop, status == 2) ~ sex,
  data = pbc,
  k = 4,
  inits = c(-12, 3, 0, 0, 0, 0, 0)
)