Skip to contents

gsm.simaerep Details

KRI.Rmd

Installation 

install.packages("pak")
pak::pak("Gilead-BioStats/clindata")
pak::pak("Gilead-BioStats/gsm.core")
pak::pak("Gilead-BioStats/gsm.mapping")
pak::pak("Gilead-BioStats/gsm.kri")
pak::pak("Gilead-BioStats/gsm.reporting")
pak::pak("IMPALA-Consortium/gsm.simaerep")

Load

{simaerep} Defaults

  • 1000 replicates of bootstrap resampling iterations
  • By default were are using the thresholds 99%, 95% and do not apply multiplicity correction
  • Events NA/NULL in numerator of denominator date columns will be dropped
  • By default {gsm.simaerep} uses all denominator events. When visits are used as denominator events filtering out unscheduled visits will result in greater snychronisation of the visit counts which will be more accurrate especially in small trials.

A higher number of replicates will produce more stable results especially for KRI based on low event counts such as treatment discontinuations of which only a small fraction of patients will have at most one event. kri0008.yaml Treatment Discontinuation provides an example on how to change the number of iterations.

Multiplicity Correction reduces the noise especially for trials with a high number of sites and makes outliers stand out more. It is the default setting for {simaerep} as results appear more plausible when looking at a single study. However, the Benjamini-Hochberg procedure eliminates small differences between the probability scores and assigns shared ranks and identical scores to some of the sites. This makes results less reproducible. On a portfolio level we have also observed a small decrease in overall statistical performance when multiplicity correction is used.

We recommend multiplicity correction if the purpose of the probability scoring is to flag sites for visually inspection on a plot displaying all sites of the study at once kri0006.yaml Treatment Discontinuation demonstrates how to switch on the multipicity correction. We advise to decrease the threshold to 95% and 75%. Increasing the number of bootstrap iterations will also help to get more reproducible results.

When dropping of event with NA date values is not acceptable relevant date imputation steps need to be added.

{gsm.simaerep} Preprocessing Options

Input_CumCount() has some more complex built-in pre-processing options that can handle special types of clinical events that surpass the simple imputation of missing dates.

Orphaned Clinical Events

Sometimes clinical events such as protocol deviations will be assigned to a site but not to a subject. {simaerep} strictly requires events to be assigned to a subject for including them in the simulations.

Input_CumCount() can treat such events in two ways either filter them (default) so that they are removed from the analysis or assign them to a random patient that is enrolled at the site at the time of the event.

However, we will only assign events that occurr between the first and the last + 30 days denominator event on site.


# we create a Numerator dataframe with missing subjet id
dfNumerator <- clindata::ctms_protdev %>%
  rename(subjid = subjectenrollmentnumber) %>%
  left_join(
    clindata::rawplus_dm %>% 
    select(subjid, siteid),
    by = "subjid"
  ) %>%
  filter(!is.na(deviationdate), !is.na(siteid)) %>%
  # set 30% of subjectid per subject to NA
  arrange(runif(n())) %>%
  mutate(rnk = row_number() / n(), .by = subjid) %>%
  mutate(subjid = ifelse(rnk < 0.3, NA, subjid))

total_events <- nrow(dfNumerator)

total_events
#> [1] 1454

linked_events <- dfNumerator %>%
  filter(! is.na(subjid)) %>%
  nrow()

linked_events
#> [1] 1188

dfCumCount <- Input_CumCount(
    dfSubjects = clindata::rawplus_dm,
    dfNumerator = dfNumerator,
    dfDenominator = clindata::rawplus_visdt %>% mutate(visit_dt = lubridate::ymd(visit_dt)),
    strSubjectCol = "subjid",
    strGroupCol = "siteid",
    strGroupLevel = "Site",
    strNumeratorDateCol = "deviationdate",
    strDenominatorDateCol = "visit_dt",
    strOrphanedMethod = "filter"
  ) %>%
  filter(Denominator == max(Denominator), .by = c(SubjectID))

filtered_events <- sum(dfCumCount$Numerator)

stopifnot(linked_events == filtered_events)

dfCumCountOrphans <- Input_CumCount(
    dfSubjects = clindata::rawplus_dm,
    dfNumerator = dfNumerator,
    dfDenominator = clindata::rawplus_visdt %>% mutate(visit_dt = lubridate::ymd(visit_dt)),
    strSubjectCol = "subjid",
    strGroupCol = "siteid",
    strGroupLevel = "Site",
    strNumeratorDateCol = "deviationdate",
    strDenominatorDateCol = "visit_dt",
    strOrphanedMethod = "assign"
  ) %>%
  filter(Denominator == max(Denominator), .by = c(SubjectID))

assigned_events <- sum(dfCumCountOrphans$Numerator)
  
assigned_events
#> [1] 1375

filtered_events / total_events
#> [1] 0.8170564

assigned_events / total_events
#> [1] 0.9456671

Terminal Binary Events

Binary Events that can only occur once per patient and mark an early drop out of the study can create a survivor bias when sampling replacement patients.

It is recommended in this case to use planned visits as the denominator instead of actual visits. However planned visits are not often available, therefore Input_CumCount() can attempt to estimate planned visits.

It will try to rank regular visit instances using the subject-level dates and determine the median time that has passed between them.

For this it will only consider visits that do not match a specific SQL like pattern and which are recorded for a minimum ratio of subjects.

Extrapolated visits that go beyond the last observed visit of patients that did not discontinue will not be generated.


dfDenominator <- clindata::rawplus_visdt %>%
    mutate(visit_dt = lubridate::ymd(visit_dt))

dfNumerator <- clindata::rawplus_studcomp %>%
  mutate(mincreated_dts = lubridate::ymd_hms(mincreated_dts))

vLikePatternInstanceName <- c("%unsch%", "%disc%")

strInstanceNameCol <- "instancename"

dfInputExtra <- Input_CumCount(
  dfSubjects = clindata::rawplus_dm,
  dfNumerator = dfNumerator,
  dfDenominator = dfDenominator,
  strGroupCol = "siteid",
  strSubjectCol = "subjid",
  strGroupLevel = "Site",
  strNumeratorDateCol = "mincreated_dts",
  strDenominatorDateCol = "visit_dt",
  vLikePatternInstanceName = vLikePatternInstanceName,
  strInstanceNameCol = strInstanceNameCol,
  nMinSubjectRatioInstance = 0.7
)

dfInputRegular <- Input_CumCount(
  dfSubjects = clindata::rawplus_dm,
  dfNumerator = dfNumerator,
  dfDenominator = dfDenominator,
  strGroupCol = "siteid",
  strSubjectCol = "subjid",
  strGroupLevel = "Site",
  strNumeratorDateCol = "mincreated_dts",
  strDenominatorDateCol = "visit_dt"
)

dfInputRegular %>%
  filter(SubjectID == "0002") %>%
  knitr::kable()
SubjectID GroupID GroupLevel Numerator Denominator
0002 76 Site 0 1
0002 76 Site 0 2
0002 76 Site 0 3
0002 76 Site 0 4
0002 76 Site 1 5 

dfInputExtra %>%
  filter(SubjectID == "0002") %>%
  knitr::kable()
SubjectID GroupID GroupLevel Numerator Denominator
0002 76 Site 0 1
0002 76 Site 0 2
0002 76 Site 0 3
0002 76 Site 1 4
0002 76 Site 1 5
0002 76 Site 1 6
0002 76 Site 1 7
0002 76 Site 1 8
0002 76 Site 1 9
0002 76 Site 1 10
0002 76 Site 1 11
0002 76 Site 1 12
0002 76 Site 1 13
0002 76 Site 1 14
0002 76 Site 1 15
0002 76 Site 1 16
0002 76 Site 1 17
0002 76 Site 1 18
0002 76 Site 1 19
0002 76 Site 1 20