Classify the census: robust, sensitive or fragile, and marginal

Applies the definitions: fragile if any counted row lies outside the base 95% confidence interval (the margin-of-error rule, the default from 0.7.0); otherwise sensitive if any counted row changes significance at alpha; otherwise robust. Marginal is a separate flag on the base p-value, true when it lies within alpha * marginal_band.

Usage

stress_verdict(
  census,
  marginal_band = c(0.5, 2),
  rule = c("interval", "interval_or_direction")
)

Arguments

census

A stress_census.

marginal_band

Multiplicative band around alpha for the marginal flag; the default flags a base p between alpha / 2 and 2 * alpha.

rule

"interval" (default): fragile means outside the base confidence interval. "interval_or_direction": outside the interval or a reversed direction (the rule of ballast 0.6 and earlier).

Value

A list of class stress_verdict: stability over every counted row, stability_sensitivity over the rows that keep the estimand (role = "sensitivity"), marginal, rule, the counts behind them, and the rows that drove the classification.

Details

For a significant base result the interval excludes zero, so a refit that reverses the direction always lies outside it and the two rules agree. They differ only for a non-significant base result: there a sign change of an estimate near zero is a small move inside the interval, and the default rule does not call it fragile. rule = "interval_or_direction" restores the 0.6 behaviour, in which any reversal is fragile.

Examples

spec <- stress_spec(mmrm::fev_data, "FEV1", "USUBJID", "AVISIT", "ARMCD", "PBO",
                    covariates = ~ RACE + SEX + FEV1_BL)
cen <- stress_census(spec, stress_covariance(spec, c("us", "cs")), stress_covariates(spec))
v <- stress_verdict(cen)
v$stability
#> [1] "robust"
v
#> Base: estimate 4.4, 95% CI 1.09 to 7.71, p = 0.00948 (alpha 0.05)
#> Counted: 5 alternative analyses, 0 deletions; 5 significant, 0 crossed alpha
#> Outside base CI: 0; direction reversed: 0; largest shift: 0.364 base SE
#> Verdict (interval rule): ROBUST over all counted rows; ROBUST over the 5 sensitivity analyses (same estimand; 0 supplementary rows)