Smallest combination of analytic changes that leaves the confidence interval

A greedy search for the fewest analytic changes, made together, that move the estimate outside the pre-specified 95% confidence interval. At each step every change from a family not yet used is tried on top of the changes already made, and the one that moves the estimate furthest in the search direction is kept. The search stops when the estimate leaves the interval or max_j changes have been made. It runs towards zero (the direction that weakens the result) and, with direction = "both", also away from zero; j is the smaller count.

Usage

stress_path(
  spec,
  changes = path_changes(spec),
  max_j = 3,
  direction = c("both", "toward_zero", "away_from_zero"),
  seed = 1,
  cores = 1L
)

Arguments

spec

A stress_spec(): the pre-specified analysis.

changes

Candidate changes from path_changes().

max_j

Largest number of changes to combine.

direction

"both", "toward_zero" or "away_from_zero".

seed

Seed for the reference-based imputations.

cores

Worker processes (or a cluster) for the candidates of a step.

Value

A stress_rows data frame with one row per step of each direction searched (family is "analytic path"), with attribute "path": a list with j (NA if not reached within max_j), reached, direction, changes (the labels of the path that reached the boundary first), assumption_dependent and the base estimate and interval.

Details

Changes combine: a reference-based or delta-adjusted imputation chosen at one step is kept, and later changes are evaluated under it. A path that leaves the interval through either is flagged assumption_dependent, so that the reader can see the path ran through a missing-data assumption rather than through the observed data alone.

Because the search is greedy, j is an upper bound on the smallest combination; to first order the shifts of different families add, so the greedy path is the worst case when they do.

The search is exact but avoids repeated work: the imputation draws are made once per imputation model and reused by every change that leaves the model alone, a state reached twice (by both directions, or by two orders of the same changes) is fitted once, and the second direction is searched only as far as it could still give a shorter path than the first.

See also

stress_fragility() for the deletion searches.

Examples

# \donttest{
d <- simulate_rare_trial(25, delta = 6, population = "mixture", seed = 2)
spec <- stress_spec(d, "y", "id", "visit", "arm", "REF", covariates = ~ base + fast,
                    method = "Satterthwaite", vcov = "Asymptotic")
p <- stress_path(spec, path_changes(spec, reference = "J2R", m = 5), max_j = 3)
attr(p, "path")$j
#> [1] NA
# }