Interpreting the verdict

Definitions

Every counted refit is scored against the pre-specified result:

• percent of the pre-specified estimate and the shift in pre-specified standard errors;
• whether it lies inside the pre-specified 95% confidence interval;
• whether it keeps the direction of effect;
• whether it is significant at alpha, and whether that differs from the pre-specified analysis.

The verdict then reads:

• fragile: some counted alternative reverses the direction or lands outside the pre-specified interval;
• sensitive: no alternative does that, but some alternative crosses alpha;
• robust: no alternative crosses alpha.

Marginal is a separate flag on the pre-specified p-value: it is set when p lies within a band around alpha (the default band is alpha/2 to 2 alpha). A marginal result is expected to be sensitive, because a p-value near 0.05 crosses 0.05 under small perturbations whatever the true effect. That is the point of keeping the two apart: “sensitive and marginal” says the trial was sized so that its p-value landed near alpha, “fragile” says the conclusion depends on a decision.

Sensitivity versus supplementary

ICH E9(R1) distinguishes sensitivity analyses, which keep the estimand and vary an assumption or a modelling choice, from supplementary analyses, which target a different estimand. Every row carries a role. Model families (covariance, inference, covariates, weighting, estimator class), deletions, re-randomisation and the imputation families are sensitivity analyses; alternative data sets, populations, other visits and intercurrent-event strategies default to supplementary. The verdict is reported over all counted rows and over the sensitivity rows alone.

The counting rule

Rows that reproduce the pre-specified analysis exactly are not counted. A specification reached by two families (the unstructured covariance appears in the covariance panel and in a grid, say) is counted once. Tipping-point grids are searches, not alternatives; the delta-zero multiple-imputation row is an alternative. Deletions are counted but summarised separately from alternative analyses, because their number is a property of the sample size, not of the analyst’s choices.

What the simulation shows

stress_calibrate() simulates trials from a planned design and runs a light stress test on each. The pattern it produces, and which the methods paper reports in full, is:

• under the null, a significant result is nearly always classed sensitive or fragile, because its p-value sits near alpha by construction;
• at high power, significant results are mostly robust, and the few sensitive ones are those whose p-value happened to land near alpha;
• “fragile” is rare under a true effect of the planned size and common under the null.

A small run (the vignette keeps it to a few trials) looks like this:

library(ballast)
stress_calibrate(n_per_arm = 30, power = c(0, 0.8), n_trials = 3, seed = 1)
#> Stress-test calibration: 30 per arm, 4 visits, SD 10, AR(1) 0.6, dropout 0.08 per visit
#>  power_target delta trials power_observed marginal robust sensitive fragile
#>           0.0  0.00      3          0.000        0  0.667     0.333       0
#>           0.8  7.33      3          0.333        0  0.667     0.333       0
#>  sensitive_given_sig fragile_k1_given_sig
#>                   NA                   NA
#>                    1                   NA

The deletion search

stress_fragility() is a greedy forward search for the smallest set of subjects whose removal changes significance, leaves the pre-specified interval, or reverses the direction. The number of steps is an analogue of the fragility index for a continuous longitudinal endpoint, with two differences: the unit is a participant, not an event, and the target can be the interval or the direction, which separates “loses significance” from “changes conclusion”. Because the search is greedy the count is an upper bound on the true minimum, and the subjects it names are a diagnostic, not a verdict on them.

Explaining a crossing

stress_explain() takes one row of the census and attributes the difference between its estimate and the pre-specified estimate to subjects, by comparing each subject’s leave-one-out influence under the two specifications. Use it on the rows that crossed alpha before writing the interpretation.