Adding lispmds-style merging to mergeMaps function

[drserajames]

Racmacs and lispmds have different merging criteria at present. When there's a mix of numeric and < titres, Racmacs converts these to numeric and takes the GMT whereas lispmds converts them all to < and takes the maximum. Additionally Racmacs uses the sample SD and lispmds uses the population SD. Lispmds does not support > titres, these are handled using a preprocessing step that converts them to numeric or *.

Proposal (using the population SD)

1. If all titres are <, keep the minimum (Racmacs & lispmds)
2. If all titres are >, keep the maximum (Racmacs)
3. Convert > and < to numeric, calculate the population SD. If greater than a threshold (default 1), set to DONTCARE (Racmacs & lispmds plus user-defined SD threshold)
4. If no < or >, calculate GMT (Racmacs & lispmds)
5. If a mix of numeric and >, calculate the GMT (Racmacs & lispmds)
6. If a mix of < and numeric, convert all to < and take the maximum (lispmds)
7. If a mix of <, numeric and >, convert > to numeric, then convert all to < and take the maximum (lispmds)

I looked in the file test-merging.R and these are my proposed changes (# is the lines I've edited, ## is new lines):

titer_merge_tests <- tibble::tribble(
  ~expected, ~sum,
  "*",    c("*"),
  "10",   c("10"),
  "<10",  c("<10"),
  ">80",  c(">80"),
  ".",    c("."),
  "10",   c(10, 10),
  "10",   c("*", "10", "10", "*"),
  "<10",  c("<10", "<40", "<20", "*"),
  ">40",  c(">10", ">40", ">20", "*"),
  "*",    c("<10", "20", "80", "*"),
  "*",    c("<10", "40", "*", "*"),    #
  "*",    c("<10", "40", "*", "."),    #
  "<40",   c("<10", "20", "10", "*"),  #
  "<40",   c("<10", "20", "10", "*"),  # duplicate row?
  "<40",   c("<10", "20", "10", "."),  #
  "*",    c("*", "*", "*", "*"),
  ".",    c(".", "."),
  "*",    c(".", "*")
  "<80",   c("<20", "20", ">20", "*")  ##
  "<80",   c("<20", "20", ">20", ".")  ##
)

And changing the sd limit function to allow a 4 fold (bottom two are the new tests)

test_that("sd_lim working", {
  
  expect_equal(
    "<40",
    ac_merge_titers(c("<10", "20"), options = RacMerge.options(sd_limit = NA))
  )
  
  expect_equal(
    "*",
    ac_merge_titers(c("<10", "20"), options = RacMerge.options(sd_limit = 0.9))
  )
  
  expect_equal(
    "<40",
    ac_merge_titers(c("<10", "20"), options = RacMerge.options(sd_limit = 1))
  )
  
  expect_equal(
    "*",
    ac_merge_titers(c(rep("<10",50), rep("20",50), "40"), options = RacMerge.options(sd_limit = 1))
  )
  
})

[drserajames]

In light of further discussion, the default behaviour will be merging all titres (no exclusions). There will be an option to have a sample SD limit when merging, with a default of 1.5. There will the options for a user defined function, plus an option for lispmds-like merging to allow backwards compatibility.

I've modified the titer_merge_tests and commented what I think each tests (# is comment, ## is lines I've edited, ### is new lines)

titer_merge_tests <- tibble::tribble(
  ~expected, ~sum,
  "*",    c("*"), 			# * unchanged
  "10",   c("10"),			# numeric unchanged
  "<10",  c("<10"),			# less than unchanged
  ">80",  c(">80"),			# greater than unchanged
  ".",    c("."),			# . unchanged
  "10",   c(10, 10),			# two same numeric, merges same
  "10",   c("*", "10", "10", "*"),	# two same numeric, two *,  merges same
  "<10",  c("<10", "<40", "<20", "*"),	# all less than go to min
  ">40",  c(">10", ">40", ">20", "*"),	# all greater than fo to max
  "160",  c("10", "160", "2560", "*"),	## very variable not set to . (default sd_limit=NA), including *
  "<40",  c("<10", "20", "*", "*"),	## very variable not set to . , including less than & *
  "<40",  c("<10", "20", "*", "."),	## very variable not set to . , including less than, * & .
  "<40",  c("<10", "20", "10", "*"),	## similar to previous, suggest delete
  "<40",  c("<10", "20", "10", "*"),	## duplicate of row above, suggest delete
  "10",   c("<10", "20", "10", "."),	## similar to previous, suggest delete
  "*",    c("*", "*", "*", "*"),	# all * merge to *
  ".",    c(".", "."),			# all . merge to .
  "*",    c(".", "*")			# mix of . & * merge to .
  "<80",  c("<20", "20", ">20", "*")  	### mix of numeric, less than, greater than & *, convert greater than and merge to < (highest log titre +1)
  "<80",  c("<20", "40", ">10", ".")  	### as previous line with numeric as the highest instead of greater than & * instead of .
)

I'm don't think "sd_lim working" needs changing any more. Instead I propose a test of the lispmds-like merging.

test_that("lispmds working", {
  
  # test n=2 allows 4-fold, but not 8-fold
  expect_equal(
    "<40",
    ac_merge_titers(c("<10", "20"), options = RacMerge.options(merge_function = "lispmds"))
  )
  
  expect_equal(
    "*",
    ac_merge_titers(c("<10", "80"), options = RacMerge.options(merge_function = "lispmds"))
  )

  # test n=102 to check the sd limit is tighter at high n
  expect_equal(
    "*",
    ac_merge_titers(c("10",rep("40", 50), rep("160",50) "20"), options = RacMerge.options(merge_function = "lispmds"))
  )

})

I don't know whether you want a variable sd_limit with lispmds merging, I can add more tests if so. Otherwise, I think there should be a warning if the user sets sd_limit and merge_function.

I can't write a test for user defined function because I still haven't got my head around what it would look like.

[shwilks]

How about this for the tests? Testing the 3 options we will have, with "conservative" being the default:

titer_merge_tests <- tibble::tribble(
  ~titers,                      ~conservative,  ~likelihood,  ~lisp,
  c("*"),                       "*",            "*",          "10",   # * unchanged
  c("10"),                      "10",           "10",         "10",   # numeric unchanged
  c("<10"),                     "<10",          "<10",        "<10",  # less than unchanged
  c(">80"),                     ">80",          ">80",        ">80",  # greater than unchanged
  c("."),                       ".",            ".",          ".",    # . unchanged
  c(10, 10),                    "10",           "10",         "10",   # two same numeric, merges same
  c("*", "10", "10", "*"),      "10",           "10",         "10",   # two same numeric, two *,  merges same
  c("<10", "<40", "<20", "*"),  "<10",          "<10",        "<10",  # all less than go to min
  c(">10", ">40", ">20", "*"),  ">40",          ">40",        ">40",  # all greater than to to max
  c("10", "160", "2560", "*"),  "160",          "160",        "*",    # very variable not set to . (default sd_limit=NA), including *
  c("<10", "20", "*", "*"),	    "<40",          "10",         "<40",  # very variable not set to . , including less than & *
  c("<10", "20", "*", "."),	    "<40",          "10",         "<40",  # very variable not set to . , including less than & *
  c("*", "*", "*", "*"),        "*",            "*",          "*",    # all * merge to *
  c(".", "."),                  ".",            ".",          ".",    # all . merge to .
  c(".", "*")                   "*",            "*",          "*",    # mix of . & * merge to *
  c("<20", "20", ">20", "*")    "<80",          "20",         "<80",  # mix of numeric, less than, greater than & *, convert greater than and merge to < (highest log titre +1)
  c("<20", "40", ">10", ".")    "<80",          "20",         "<80",  # as previous line with numeric as the highest instead of greater than & * instead of .
)

Am not sure what to do with the > values in the "lispmds" merge option though?

[drserajames]

A couple of thoughts

• Should be a "*" for lisp on that first row
• Should c("10", "160", "2560", "*") merge to "." not "*" for lisp? Similarly c(".", "*") should merge to ".". Or perhaps I'm missing something?
• I wouldn't say c("<10", "20", "*", "*") & c("<10", "20", "*", ".") is very variable. They are tests of mix of numeric & less than

I think for lisp, the > should be converted to numeric because that's the most likely pre-processing I would do with lisp.

c("10", ">20", "*", "."),	    "20",          "20",         "20",

[shwilks]

Thanks, for the last 2 cases we used to replace with "*" if there was a mix of < and >, but did we simply decide not to do that any more? I guess so?

[drserajames]

I don't remember the discussion.

I thought Racmacs converted all to numeric took the GMT if SD<1 (that's what I put in the summary document so I'll need to correct it). lispmds would have pre-processed the > to numeric and so would treat it as a case of < & numerics. I think either is fine because practically speaking for flu it's likely to be <10 and >1280 so the SD limit would make it * anyway.

[shwilks]

In the current version Racmacs has a special rule to set any mix of < and > to *, which mirrors what acmacs did. We agreed the default sd would be NA so without that rule <10 and >1280 would become <2560 for the new default "conservative" method and around 113 for the "likelihood" method.

I suggest we keep the old Racmacs / acmacs rule and they become * with a warning.

[drserajames]

I agree to with merging < & > to *. I can't remember what Derek said.

<10 and >1280 would actually merge to <5120.

[shwilks]

ok, and yes true re <5120

[shwilks]

I just pushed version 1.2.0 to shwilks/Racmacs which implements these merge changes, although no merge messages or warnings about it yet.

[drserajames]

Thanks!