Feature Request: ard_summary but for hierarchies

[kieranjmartin]

What's the feature?

At the moment, ard_summary performs some function in every possible subgroup of each by variable or in every subgroup with observations in strata.

However, what it cannot mimic is the hierarchical behaviour, where we have every combination of a plausible subgroup, where each variable represents a level of hierarchy, which is what we need for AEs. Unfortunately, at the moment ard_hierarchical is hard wired to tabulate: calculate N.

I have come across at least one situation where this would be useful: Exposure-adjusted incidence rate (EAIR) requires calculating total patient time at risk, which is dependent on the sub category you are calculating for. So for covid I need to know when those patients experienced their first covid exposure, and then calculate the risk from the remainder of the data.

Now technically I could use mvr_summary here, but the total size of the by combinations (most of which are nonsensical) means that R actually doesn't even get to any calculations; it hangs generating all of the possible combinations!

I think the best solution is something like ard_hierarchical_summary to allow for this situation

[ddsjoberg]

HI @kieranjmartin !

I am not sure I follow 100%. The hierarchical summary uses ard_tabulate(strata) in the background to give us an ARD of observed combinations.

[kieranjmartin]

So lets say I have some function, exposure_fun which I want to apply to my data using ard_mvsummary, but I want to use the hierarchical structure used by ard_hierarchical_summary. I can sort of replicate it using

    cards::ard_mvsummary(
      statistic = everything()  ~ list(exposure = expfun),
      variables = trtsdt,
      by = ARM 
      strata = c(lvl01, lvl02, lvl03, lvl04)
    )

But this only gives me the bottom level; I also want the stat for each level I need something like

  cards::ard_stack(
    data = indata,
    .by = ARM,
    cards::ard_mvsummary(
      statistic = everything() ~ list(exposure = expfun),
      variables = trtsdt,
      strata = lvl01
    ),
    cards::ard_mvsummary(
      statistic = everything()  ~ list(exposure = expfun),
      variables = trtsdt,
      strata = c(lvl01, lvl02)
    ),
    cards::ard_mvsummary(
      statistic = everything()  ~ list(exposure = expfun),
      variables = trtsdt,
      strata = c(lvl01, lvl02, lvl03)
    ),
    cards::ard_mvsummary(
      statistic = everything()  ~ list(exposure = expfun),
      variables = trtsdt,
      strata = c(lvl01, lvl02, lvl03, lvl04)
    )
    
  )  

(maybe you can do something clever with map here to avoid the repetition, I'm not actually sure: the implementation in ard_stack_hierarchical doesn't actually use ard_stack and instead uses append with a loop)

My point here is this is not terribly intuitive. I'd like to be able to use the hierarchical functions here but they currently expect a denominator and will only calculate counts. So my thought was probably a hierarchical version of mvsummary might be good

[ddsjoberg]

Ahh, I see, thanks for the calcification. I would approach it like this:

vars <- c("lvl01", "lvl02", "lvl03", "lvl04")

map(
  seq_along(vars),
  ~ cards::ard_mvsummary(
      statistic = everything()  ~ list(exposure = expfun),
      variables = trtsdt,
      by = ARM 
      strata = all_of(vars[1:.x])
   )
) |> 
  cards::bind_ard()

I haven't found this pattern to be common, but if others are also coming across it frequently we could include some kind of helper for it.