Exact Paired Sign-Flip Tests of Within-Subject Differences
Source:R/analysis-paired.R
ExactPairedTests.RdFor each declared group of rows, tests whether paired differences are symmetric about zero with the exact Wilcoxon signed-rank test over all \(2^n\) sign assignments, together with the exact two-sided sign test and, optionally, a Clopper-Pearson interval for the fraction of subjects whose difference lies in an expected direction. Leave-one-subject-out results can be requested instead of the full tests.
Usage
ExactPairedTests(
data,
value,
groups = NULL,
id = NULL,
expected = NULL,
family = NULL,
omit = c("none", "each"),
conf_level = 0.95,
max_n = 500L
)Arguments
- data
Data frame in long form: one row per subject and test.
- value
Name of the numeric difference column (for a one-sample test, the values themselves).
- groups
Optional names of columns that define separate tests.
NULLtreats all rows as one test. Output groups follow first appearance.- id
Optional subject column; required for
omit = "each". Subjects must be unique within a group.- expected
Optional expected direction:
"less"or"greater"for every group, or the name of a column holding"less"/"greater"(constant within a group). Adds the number of subjects in that direction and its Clopper-Pearson interval.- family
Optional columns defining multiplicity families; adds
family_id(values joined by" :: ").hypothesis_idjoins the group values the same way. Both are ready forAdjustPatientFamilies().- omit
"none"(full tests) or"each": one row per group and omitted subject (all subjects of the group, in data order), with the omitted subject in columnomitted.- conf_level
Confidence level of the direction interval.
- max_n
Largest number of nonzero differences accepted.
Value
A data frame with the group columns, n_available, n_nonzero,
n_zero, positive, negative, median_difference and
mean_difference (of the nonzero differences), signed_rank_sum,
p_value (signed-rank), sign_p_value, status, optionally expected,
expected_direction, direction_ci_low, direction_ci_high,
conf_level, and hypothesis_id (plus family_id).
Details
Non-finite differences are unavailable; zero differences are excluded from the signed-rank and sign tests (they still count as available subjects and as not being in the expected direction). Ranks of absolute differences use midranks for ties. The two-sided p value is the fraction of sign assignments whose absolute signed-rank sum is at least the observed one. It is computed by dynamic programming over twice the midranks (integers), so it equals full enumeration exactly for up to 52 nonzero differences without enumerating them.
Examples
d <- data.frame(test = rep(c("a", "b"), each = 6), subject = rep(1:6, 2),
difference = c(-3, -1, -2, -5, 1, -4, 2, -1, 3, 0, 1, 2))
ExactPairedTests(d, "difference", groups = "test", id = "subject",
expected = "less")
#> test n_available n_nonzero n_zero positive negative median_difference
#> 1 a 6 6 0 1 5 -2.5
#> 2 b 6 5 1 4 1 2.0
#> mean_difference signed_rank_sum p_value sign_p_value status expected
#> 1 -2.333333 -18 0.09375 0.21875 EXACT less
#> 2 1.400000 12 0.18750 0.37500 EXACT less
#> expected_direction direction_ci_low direction_ci_high conf_level
#> 1 5 0.358765421 0.9957893 0.95
#> 2 1 0.004210745 0.6412346 0.95
#> hypothesis_id
#> 1 a
#> 2 b
ExactPairedTests(d, "difference", groups = "test", id = "subject",
omit = "each")
#> test omitted n_available n_nonzero n_zero positive negative
#> 1 a 1 5 5 0 1 4
#> 2 a 2 5 5 0 1 4
#> 3 a 3 5 5 0 1 4
#> 4 a 4 5 5 0 1 4
#> 5 a 5 5 5 0 0 5
#> 6 a 6 5 5 0 1 4
#> 7 b 1 5 4 1 3 1
#> 8 b 2 5 4 1 4 0
#> 9 b 3 5 4 1 3 1
#> 10 b 4 5 5 0 4 1
#> 11 b 5 5 4 1 3 1
#> 12 b 6 5 4 1 3 1
#> median_difference mean_difference signed_rank_sum p_value sign_p_value
#> 1 -2.0 -2.20 -12 0.1875 0.3750
#> 2 -3.0 -2.60 -13 0.1250 0.3750
#> 3 -3.0 -2.40 -12 0.1875 0.3750
#> 4 -2.0 -1.80 -12 0.1875 0.3750
#> 5 -3.0 -3.00 -15 0.0625 0.0625
#> 6 -2.0 -2.00 -12 0.1875 0.3750
#> 7 1.5 1.25 7 0.3750 0.6250
#> 8 2.0 2.00 10 0.1250 0.1250
#> 9 1.5 1.00 7 0.3750 0.6250
#> 10 2.0 1.40 12 0.1875 0.3750
#> 11 2.0 1.50 8 0.2500 0.6250
#> 12 1.5 1.25 7 0.3750 0.6250
#> status hypothesis_id
#> 1 EXACT a :: 1
#> 2 EXACT a :: 2
#> 3 EXACT a :: 3
#> 4 EXACT a :: 4
#> 5 EXACT a :: 5
#> 6 EXACT a :: 6
#> 7 EXACT b :: 1
#> 8 EXACT b :: 2
#> 9 EXACT b :: 3
#> 10 EXACT b :: 4
#> 11 EXACT b :: 5
#> 12 EXACT b :: 6