Using a date with multiple values when defining population

[elsie-h]

I'm working on a project in which I want to select individuals who have not been vaccinated in the 12 weeks following their eligibility date (project not yet approved so in my personal GitHub repo - can give access on request) . Therefore, when selecting the population I want patients.registered_as_of(elig_date), and let's assume there are 10 different possible values for elig_date. I'll also want to use elig_date when defining other variables in the study definition.

I've allowed for this by creating a different study definition for each elig_group, which was done by adapting @wjchulme's create-project.R such that the study definitions are created from a file called template_study_definition.py:

# define number of eligibility groups
n_elig_groups <- 10

create_study_definitions<- 
  str_c("#!/bin/sh",
        "# Run this script to create a study_definition for each eligibility date",
        "",
        str_c("for i in {1..",n_elig_groups,"}; do"),
        "sed -e \"s;%placeholder_elig_date%;elig_date_$i;g\" -e \"s;%placeholder_elig_grp%;$i;g\" ./analysis/template_study_definition.py > ./analysis/study_definition_$i.py;",
        "done;", sep = "\n")

create_study_definitions %>%
  writeLines(here::here("analysis/create_study_definitions.sh"))

# create study definitions from template_study_definition.py
check_create <- try(processx::run(command="bash", args= "analysis/create_study_definitions.sh"))

if (class(check_create)=="try-error") stop("Study definitions not created.")

Then I can create the actions for the project.yaml using the action function defined in create-project.R.

stud_def_actions <-  lapply(1:n_elig_groups,
                            function(x)
                              action(
                                name = str_c("elig_grp_", x),
                                run = str_c("cohortextractor:latest generate_cohort --study-definition study_definition_", x),
                                needs = list("design"),
                                highly_sensitive = list(
                                  cohort = str_c("output/input_",x,".csv")
                                )
                              ))

I was chatting to @wjchulme about this in case he knew of a more elegant way to do this, and he recommended posting it on here to see if anyone else has as a better solution/can use my solution if it's useful!

(I appreciate this might be a bit difficult to follow without access to the repo so please let me know and I can give you access!)

[wjchulme]

Worth pointing out that one failed attempt at solving this neatly was to use patients.categorised_as to create the group-specific date variable:

population = patients.registered_as_of("elig_date"),
jcvi_group = patients.categorised_as(
   {
      "group 1": " age>=80 OR carehomeresident OR carehomestaff",
      "group 2": "healthcareworker",
      "group 3": "age>=75",
      ...
   }
),
elig_date = patients.categorised_as(
   {
      "2020-12-08": "jcvi_group = 'group 1'",
      "2020-12-08": "jcvi_group = 'group 2'",
      "2021-01-04": "jcvi_group = 'group 3'",
      ...
   }
),

It doesn't work because patient.categorised_as returns categoricals, which won't be implicitly interpreted as dates.

[evansd]

@wjchulme's proposal looks to me like The Right Way of solving the problem except, as you point out, it doesn't work for very annoying typing reasons. This is, yet again, one of those issues that cohortextractor-v2 was designed to avoid but that doesn't help you much in the here and now.

However, it might be possible to tweak the categorised_as function a bit to support returning different types. Off the top of my head I can't decide whether that would be really quite straightforward or hideously difficult so I'll have to look into it more. But it feels worth at least some initial investigation as if it is possible it might provide an elegant solution to quite a few problems like this.

[elsie-h]

Thanks for looking into this, @evansd. As I'm new to working with OpenSAFELY I've not run a study definition yet on the TPP data, but I'm concerned that the workaround that I posted will be painfully slow as I currently have 22 eligibility groups.

[evansd]

I've fixed the problem with using dates in categorised_as() so something like @wjchulme's proposed solution should now work:
opensafely-core/cohort-extractor#656

One minor thing in passing though: you'll need to take the spaces out of the category names and use e.g. group_1 rather than group 1 or it will complain when you try to use them in an expression.