Remove or shorten the subjects flag_identifiable_subjects() flags

Applies a remediation policy to the outliers found by flag_identifiable_subjects(). By default a subject flagged only for an unusually long follow-up is truncated back to the cohort's longest ordinary follow-up (its late rows dropped), and a subject flagged for any other reason – an extreme DV, a rare dose level, or an unusual dose count – is dropped entirely.

Usage

remediate_identifiable_subjects(
  data,
  roles,
  source = NULL,
  time = c("truncate", "drop", "keep"),
  other = c("drop", "keep"),
  threshold = 3.5,
  seed = NULL,
  max_tries = 20L
)

Arguments

data

A PMX dataset, typically the synthetic output.

roles

Explicit roles from pmx_roles().

source

Optional source PMX data. When given, dropped subjects are replaced by fresh avatars generated from it, so the cohort size is preserved. When NULL (default), dropped subjects are simply removed.

time

Action for a subject whose only outlier axis is follow-up time: "truncate" (default) to shorten it to the longest ordinary follow-up, "drop" to remove it, or "keep" to leave it.

other

Action for a subject flagged for any non-time reason: "drop" (default) or "keep".

threshold

Passed to flag_identifiable_subjects().

seed

Reproducibility seed for the replacement generation. The caller's random-number state is restored by synpmx_avatar().

max_tries

Maximum regeneration batches when refilling dropped subjects.

Value

data with the policy applied, carrying attributes dropped, truncated (affected subject ids), replaced (count refilled), and horizon (the follow-up truncation used, or NA).

Details

The split is deliberate. Truncation is offered only for an unusually long follow-up, the one structural outlier a value-level edit can genuinely fix: shortening a long timeline leaves a shorter but ordinary subject. Everything else is dropped – an unusually short follow-up has nothing to trim, an extreme-DV subject is elevated across its whole trajectory so removing points would only mangle it, and a rare dose cannot be trimmed without breaking the regimen. A subject flagged for both a long follow-up and another reason is also dropped, since truncation would not resolve the other reason.

When source is supplied, each dropped subject is replaced: fresh avatars are generated from source, screened by the same policy, and appended (with new ids) until the cohort is back to its original size. So the output keeps the same number of subjects, minus any it could not refill within max_tries. Truncation keeps its subject, so it never triggers a replacement.

Detection is per subject, so one long-followed patient is truncated once and one extreme patient dropped-and-replaced once – there is no row-level outlier spray. With replacement, this is a self-contained alternative to preventing structural outliers at generation time (skeleton sampling).

See also

flag_identifiable_subjects().

Examples

data <- pmx_simulated_fixture(30)
roles <- pmx_roles(
  id = "ID", time = "TIME", dv = "DV", amt = "AMT", evid = "EVID",
  cmt = "CMT", dvid = "DVID", covariates = "WT"
)
synthetic <- suppressWarnings(synpmx_avatar(data, roles, seed = 1))
#> synpmx_avatar(): dropped 9 undeclared column(s): NTIME, TAD, OCC, RATE, MDV, CENS, LIMIT, AGE, SEX.
#>   Declare a column in `keep` to carry it through verbatim.
cleaned <- remediate_identifiable_subjects(synthetic, roles, source = data)
#> remediate_identifiable_subjects(): dropped 0, truncated 0.