Per discussion in the linked issue, here is an example of a KM curve next to ggsurvplot (who I trust have implemented this correctly for KM curves).

devtools::load_all("~/packages/ggplot2/")
#> ℹ Loading ggplot2
library(survival)
model <- survfit(Surv(time, status) ~ x, data = aml)

df <- broom::tidy(model)

ggplot(df, aes(time, estimate, ymin = conf.low, ymax = conf.high)) +
  geom_step(aes(colour = strata)) +
  geom_ribbon(
    aes(fill = strata), alpha = 0.2,
    position = "connect"
  )

ggsurvfit::ggsurvfit(model) +
  ggsurvfit::add_confidence_interval() +
  theme_gray()
#> Warning: Ignoring unknown labels:
#> • `linetype = "NULL"`

^{Created on 2024-12-11 with reprex v2.1.1}

I suspect ggsurvfit does some extra steps to ensure series start at 'time = 0' and 'probability = 1', but the CI ribbons look not too bad.

I'm a bit on the fence with this, tbh. There is obvious utility in its generality, but Position adjustments shouldn't make up new data - it is only for moving existing data around.

Now, you could argue that a straight line is an infinite set of points between two extremes so it is just moving those points around, but I think that would be stretching it a bit.

Conceptually this feels like it belongs in a stat, but that would preclude it from being used together with other stats like in the examples above. However, I feel we did touch on the idea of having a stat_stack()at some point which would remove this downside

That makes sense to me. Once we can chain stats, we can reshape this into a stat_connect() or something, so I'll mark this as a draft for now.