Feature request: support of svykm objects

[larmarange]

When dealing with complex and weighted datasets, survey:svykm() should be used instead of survival::survfit(). Would it be relevant to include support of such objects?

[ddsjoberg]

Hey @larmarange !! I think it can be supported here.

Just like in gtsummay, where we've been developing an ARD-first, I plan on updating ggsurvfit to also support ARD first. Once that is done, we'd just need an ARD function for survey:svykm() and then it can simply be incorporated into the pipeline.

I don't have a timeline for this at the moment (more pressing matters to deal with at the moment), unfortunately. Also, once we generalize to accept ARD inputs, I wonder if it'll be best to support survey directly here or in a spinoff package that takes advantage of this infrastructure. Something we can decide/discuss when we get to that point. (But not sure when we will be to that point).

[larmarange]

Hi @ddsjoberg

Thanks for your feedback. I also noted that gtsummary::tbl_survfit() is not compatible with svykm().

Would a first step be to add a ard_survey_svykm() function in cardx?

[ddsjoberg]

Yeah that would be among the first for sure. The BIG item is going to be updates in ggsurvfit to handle ARDs.

Regarding ARDs, I still need to land on a consistent method for reporting variable-level and model-level statistics and how to link them. This applies to regression models, where coefs are associated with a variables in the model, and we have model levels stats like AIC (and many others).

Anyway, what i want to say is that if you write an cardx::ard_survey_svykm() function now, we'll probably need to update it in future once we work out some storage details.

[larmarange]

Let me know if you think it's too early for cardx::ard_survey_svykm(). There is no emergency. But in that case it could be relevant to have an issue open as a reminder.

So far, I'm teaching to my students the classic approach for weighted KM, and I give them a small function to get the data by time points: https://larmarange.github.io/guide-R/analyses_avancees/analyse-survie.html#analyse-de-survie-pond%C3%A9r%C3%A9e

But on a longer term, it would be nice to have a unified way to do it regardless it is weighted or not

[jinseob2kim]

How about https://github.com/jinseob2kim/jskm ?

[larmarange]

Thanks @jinseob2kim I have added in my teaching reference to jskm.

@ddsjoberg Just as a reminder when developing support of svykm() in cardx, I have drafted two exploratory functions to get times and probs from such object.

svykm_probs <- function(x,
                        probs = c(1, .75, 5, .25),
                        ci_level = .95,
                        strata = NULL) {
  if (inherits(x, "svykm")) {
    if (is.null(ci_level) | is.null(x$varlog)) {
      res <- quantile(x, probs, ci = FALSE) |> 
        dplyr::as_tibble(rownames = "prob")
    } else {
      tmp <- quantile(
        x,
        probs,
        ci = TRUE,
        level = ci_level
      )
      ci <- attr(tmp, "ci") |> 
        dplyr::as_tibble(rownames = "prob") |> 
        dplyr::rename(conf.low = 2, conf.high = 3)
      res <- tmp |> 
        dplyr::as_tibble(rownames = "prob") |> 
        dplyr::left_join(ci, by = "prob")
    }
    if (!is.null(strata))
      res$strata <- strata
    res
  } else {
    x |> 
      seq_along() |> 
      lapply(
        \(i) {
          svykm_probs(
            x[[i]],
            probs = probs,
            ci_level = ci_level,
            strata = names(x)[[i]]
          )
        }
      ) |> 
      dplyr::bind_rows()
  }
}
svykm_times <- function(x,
                        times,
                        ci_level = .95,
                        strata = NULL) {
  if (inherits(x, "svykm")) {
    idx <- sapply(
      times,
      function(t) max(which(x$time <= t))
    )
    if (is.null(ci_level) | is.null(x$varlog)) {
      res <- dplyr::tibble(
        time = times,
        value = x$surv[idx]
      )
    } else {
      ci <- confint(x, parm = times, level = ci_level)
      res <- dplyr::tibble(
        time = times,
        value = x$surv[idx],
        conf.low = ci[, 1],
        conf.high = ci[, 2]
      )
    }  
    if (!is.null(strata))
      res$strata <- strata
    res
  } else {
    x |> 
      seq_along() |> 
      lapply(
        \(i) {
          svykm_times(
            x[[i]],
            times = times,
            ci_level = ci_level,
            strata = names(x)[[i]]
          )
        }
      ) |> 
      dplyr::bind_rows()
  }
}

Some examples of use here: https://larmarange.github.io/guide-R/analyses_avancees/analyse-survie.html#courbes-de-kaplan-meier