Skip to contents

Which strata can mask their own avatars

Source: R/compare.R
unmaskable_strata.Rd

One row per declared stratum level, counting the patients in it that no avatar can safely be built on. An avatar is anchored on one real patient and copies that patient's dose times verbatim, so an arm has to contain somebody who can be masked; re-anchoring picks a different anchor, but never one from another arm, because the anchor carries its strata values into the output and moving it would silently rewrite the arm sizes.

Usage

unmaskable_strata(data, roles, min_pattern_share = 2L, coarsen_time = TRUE)

Arguments

data

A PMX dataset – the source, before generating.

roles

Explicit roles from pmx_roles().

min_pattern_share

The floor synpmx_avatar() will be run with: how many patients must share a pattern before it may be reused.

coarsen_time

Score the coarsened visit grid (TRUE, the default, and what synpmx_avatar() does) or the recorded times as given.

Value

A pmx_unmaskable_strata data frame, worst arm first, with stratum, patients, unmaskable_dosing, unmaskable_visits and safe_anchors. One row named all when the roles declare no strata.

Details

safe_anchors is therefore the number that matters. An arm with none of them will emit avatars carrying a real patient's schedule whatever the generator does, which is what synpmx_scorecard() reports as B1a and B1b, and the fix is to the study description rather than to the run – see the alert synpmx_avatar() raises.

Two ways a patient cannot be masked, counted separately because the remedies differ:

  • unmaskable_dosing – their set of dose times is shared by fewer than min_pattern_share patients, and no prefix of it can be given to an avatar either: every shorter opening is one that either a single patient stopped at, or that fewer than min_pattern_share patients passed through. Dose events are copied from the anchor as they stand, since resampling them would emit regimens the protocol never permitted, so there is nothing to substitute.

  • unmaskable_visits – the set of visits they have observations at is shared by fewer than min_pattern_share patients, and their schedule group holds no set that is shared widely enough to put in its place. Usually fixable: a wider pool, or strata = NULL so the arms share one.

This reads real patient data and is marked "restricted_not_releasable".

See also

skeleton_uniqueness() for the same question per patient, synpmx_scorecard(), which reports this as rows B1a and B1b.

Examples

data <- pmx_simulated_fixture(20)
data$ARM <- ifelse(as.integer(data$ID) %% 2L == 0L, "A", "B")
roles <- pmx_roles(
  id = "ID", time = "TIME", dv = "DV", amt = "AMT", evid = "EVID",
  cmt = "CMT", dvid = "DVID", covariates = "WT", strata = "ARM"
)
unmaskable_strata(data, roles)
#> Patients no avatar can safely be built on, by arm
#> 
#>  stratum patients unmaskable_dosing unmaskable_visits safe_anchors
#>        B       10                 0                 0           10
#>        A       10                 0                 0           10
#> 
#> Every arm can mask its own avatars.