Statistical Methods
Use an Exact Sign Test for Nonzero Paired Changes
Define the task
Define a paired difference and its direction before looking at results. A sign test asks whether positive and negative nonzero differences are equally likely; it discards magnitudes. Use one independent difference per unit and preserve the pairing rule. Zero differences are omitted by the convention used here and must be reported. This conditional sign test is not automatically a test of an unconditional population median when there is substantial probability mass at zero.
Carry out the check
Count positive and negative differences, remove exact zeros and compare the positive count with a Binomial(n,0.5) reference. The tool uses a two-sided exact probability, twice the smaller inclusive tail capped at one; for a symmetric 0.5 null this agrees with the probability-ordered exact binomial calculation. Missing and malformed entries are rejected rather than converted to zero. All-zero data leave no testable signs. Do not select a one-sided direction after inspection, and report effect size and measurement context separately because the p-value ignores how large changes are.
Worked hypothetical example
For fictional differences (1,2,3,4,−1), four of five nonzero signs are positive. The smaller binomial tail has probability (1+5)/32=0.1875; the two-sided p-value is 0.375. With five positive signs the two-sided probability is 2/32=0.0625. These examples show the discreteness of small-sample inference. A nonsignificant result does not demonstrate equivalence or absence of a meaningful paired change.
Method checked 18 November 2026 against Primary method documentation: binomtest.