"initial value in 'vmmin' is not finite" error for flexsurvspline

[Erikwjoh]

When I run flexsurvspline on the attached dataset (flexsurvspline(Surv(time,dead)~1,data=input_df,k=2)), I get the following error for any k>1: Error in optim(method = "BFGS", par = c(gamma0 = 0, gamma1 = 0, gamma2 = 0, : initial value in 'vmmin' is not finite

From reading other posts with similar issues, I gather this is likely an issue with the inits parameter, and that I might have to generate my own inits function for this particular dataset. But I haven't found any guidance on how to do this for spline fits or what goes into that parameter. Am I correct that this is what's causing the error, and if so how should an inits function be determined?

input_df: input_df.xlsx

Thanks in advance!

[chjackson]

If a model doesn't work, it's best to simplify it until it does work, and then build it up bit by bit. In this case, you could use the parameters from a fitted model with one internal knot to initialise the two-knot model. But it doesn't converge in this case. This is very small dataset with only 12 distinct event times, so I wouldn't be surprised if this (four-parameter) model were not identifiable.

fl1 <- flexsurvspline(Surv(time,dead)~1,data=input_df,k=1)
plot(survfit(Surv(time, dead) ~ 1, data=input_df))
lines(fl1, t=seq(0,10,1))
flexsurvspline(Surv(time,dead)~1,data=input_df,k=2,
               inits = c(coef(fl1), 0))

[Erikwjoh]

Many thanks for the explanation @chjackson! I found that I am able to fit a 2-knot model when I specify my own internal knot values. Is there any disadvantage to doing this? I'm new to using splines and don't have any experience in using custom knot locations.

Thanks again!

[chjackson]

The exact location of the knots doesn't usually make a big difference, as I've found.