Skip to contents

Smallest set of subjects whose removal changes the conclusion

Source: R/fragility.R
stress_fragility.Rd

Two searches for the fewest subjects whose removal moves the result to the target. The number needed is a continuous-outcome analogue of the fragility index; both searches give an upper bound on the true minimum.

Usage

stress_fragility(
  spec,
  target = c("significance", "ci", "direction"),
  max_k = 5,
  candidates = NULL,
  cores = 1L,
  method = c("greedy", "ranked")
)

Arguments

spec

A stress_spec().

target

What has to change: "significance" (the p-value crosses alpha in either direction, depending on the base result), "ci" (the estimate leaves the base confidence interval) or "direction" (the estimate changes sign).

max_k

Maximum number of subjects to remove.

candidates

Optional character vector of subject identifiers to restrict the search to (one arm, say).

cores

Number of worker processes; 1 fits sequentially.

method

"greedy" or "ranked".

Value

A stress_rows data frame with one row per step (kind = "search", so the rows are not counted in a census, since the search crosses the target by construction; param = number removed, label = the subjects removed so far). attr(, "fragility") is a list with target, method, k (the step at which the target was reached, NA if never), subjects and reached.

Details

  • method = "greedy" (default): at each step every remaining subject is removed in turn from the current data, the pre-specified model is refitted, and the subject whose removal moves the result furthest towards the target is dropped for good. About k * n refits.

  • method = "ranked": subjects are ranked once, by how far removing each alone moves the result towards the target (the leave-one-out refits), and the top 1, 2, ..., k are removed together, one refit per step. About refits, so it is practical for large trials. influences at once.

Examples

small <- mmrm::fev_data[mmrm::fev_data$USUBJID %in% paste0("PT", 1:60), ]
spec <- stress_spec(small, "FEV1", "USUBJID", "AVISIT", "ARMCD", "PBO", covariates = ~ FEV1_BL,
                    covariance = "cs", method = "Satterthwaite", vcov = "Asymptotic")
fr <- stress_fragility(spec, max_k = 1)
attr(fr, "fragility")
#> $target
#> [1] "significance"
#> 
#> $method
#> [1] "greedy"
#> 
#> $k
#> [1] 1
#> 
#> $subjects
#> [1] "PT1"
#> 
#> $reached
#> [1] TRUE
#> 
#> $base_p
#> [1] 0.1099548
#> 
#> $base_estimate
#> [1] 3.478182
#> 
attr(stress_fragility(spec, "ci", max_k = 10, method = "ranked"), "fragility")$k
#> [1] 7