Skip to contents

Enumerates every allocation with the declared group sizes. The statistic is the mean in A minus the mean in B; the two-sided p value is the fraction with absolute statistic at least the observed absolute statistic. Observed and allocated statistics are computed alike from values centred on their mean. Two statistics count as tied when they differ by no more than the rounding of this arithmetic (relative to the largest centred value) plus the rounding of the stored values (half a unit in the last place of the largest absolute value), so p values do not depend on the scale of the values. A shift leaves them unchanged while the rounding of the shifted values stays below the spacing of distinct statistics; for example, whole numbers from 0 to 20 for up to 12 patients keep their exact p values up to a shift of 1e13, while larger shifts can merge distinct statistics into ties. If an intermediate tolerance bound overflows, the same bound is evaluated with machine precision multiplied first; finite ordinary bounds are unchanged. This requires exchangeability under the null. Observational confounding, paired designs and stratified randomisation are not addressed by this test. Repeated images must be aggregated to one endpoint per patient beforehand.

Usage

ExactPatientTests(
  endpoints,
  design,
  hypotheses,
  missingness = "require_complete",
  max_allocations = 1000000L
)

Arguments

endpoints

Data frame with character endpoint_id, patient_id, dimensionless numeric vector value, and dimensionless logical vector eligible. At most one row per endpoint and patient. NA values or eligibility mark unavailable observations; NaN and infinite values are rejected. Extra columns are ignored.

design

Data frame with character contrast_id, patient_id, and group (exactly "A" or "B"). Each patient occurs once per contrast; both groups must be nonempty. Patients can belong to multiple contrasts.

hypotheses

Data frame with unique character hypothesis_id and character endpoint_id, contrast_id, family_id. Every declared row, including unavailable tests, remains in the result and family count.

missingness

Only "require_complete" is supported. All expected patients must have finite eligible endpoint values. No patient is dropped.

max_allocations

Positive integer ceiling on exhaustive allocations, at most one million. Exceeding it fails rather than using an approximation.

Value

A data frame with the registry, expected/evaluable group sizes, mean_difference, p_value, allocation count, probability grid step, and status. Grid step is a resolution bound, not a guaranteed attainable minimum two-sided p value. Unavailable effects and p values remain NA.

Examples

endpoints <- data.frame(endpoint_id = "e", patient_id = letters[1:4],
                        value = 1:4, eligible = TRUE)
design <- data.frame(contrast_id = "c", patient_id = letters[1:4],
                     group = c("A", "A", "B", "B"))
hypotheses <- data.frame(hypothesis_id = "h", endpoint_id = "e",
                         contrast_id = "c", family_id = "f")
ExactPatientTests(endpoints, design, hypotheses)
#>   hypothesis_id endpoint_id contrast_id family_id n_a_expected n_b_expected
#> 1             h           e           c         f            2            2
#>   n_a_evaluable n_b_evaluable mean_difference   p_value allocations
#> 1             2             2              -2 0.3333333           6
#>   probability_grid_step status missingness_policy inference_unit
#> 1             0.1666667  EXACT   require_complete        PATIENT