Option for weighted/unweighted `n.risk`, `n.censor`, etc. with weighted `survfit()` call

[ddsjoberg]

Thank you @therneau for the note! "When there are case weights, do we want the weighted number at risk/events/censored or unweighted? I lean towards the unweighted for labels, weighted for follow-on computaions, but could argue both sides"

[VenusSo]

@ddsjoberg Yes, unweighted is preferred for the plot output., although the best is to have an option. If it's weighted, it should to rounded to integers. Currently, it's showing weighted with a lot of decimal places. It would be great if the default can be the unweighted counts. Maybe add a fit = option in add_risktable() so that people can put weighted or unweighted fit from survfit2(). Thanks for the nice package!

[VenusSo]

Hi @ddsjoberg, attached is a minimal set of sample data and the resulting KMplot with the weights argument turned on in the following minimal function. Some of the group_1 subjects have <1 weight leading to decimal counts. My current workaround is to round off the n.risk count with the kmfit$n.risk <- round(kmfit$n.risk) line in my function.

ggsurv <- function(OSdf, weights=NA) {

  form <- as.formula(Surv(timetoevent, status) ~ drug)

  if(is.na(weights)==TRUE) {
    kmfit <- survfit2(form, data = OSdf)
  } else {
    kmfit <- survfit2(form, data = OSdf, weights = OSdf[ , weights])
  }
  
  # # round the n.risk (for weighted data)
  # kmfit$n.risk <- round(kmfit$n.risk)
  
  plot <- kmfit |>
    ggsurvfit() +
    # add_censor_mark(size = 1, alpha = 0.5) +  # lead to error in `tidyr::uncount()` for weighted data
    add_risktable(
      hjust = 0,
      risktable_stats = c("n.risk"))
  
  return(plot)
}

ggsurv(sampledata, weights="weight")

At a second thought, I think rounding the weighted count is the preferred default since this is how it's handled in the survminer::ggsurvplot(). It will be nice to provide an option to output unweighted count, for example through a a fit = option in add_risktable(). Just a suggestion. Thanks again!

[ddsjoberg]

Hi @VenusSo

You can round the stats in the risktable using glue::glue() syntax to inject a rounding function. Please review the ?add_risktable() help file for an example that calls out weighted models for this particular use case.

You can also report unweighted counts for a weighted model. Here's an example.
#235

[VenusSo]

@ddsjoberg Ah great! Thanks for pointing that out!