One-Sided Test Rationale Lock

A one-sided test is only valid when the direction of the effect was predicted before data collection. Choosing one-sided testing after seeing the data —because one direction happens to reach p<0.05 while the two-sided result does not —is a form of p-hacking that halves the nominal p-value without a corresponding reduction in false positive risk.

STEP 1 —The Pitfall

Two-sided p=0.06 can become one-sided p=0.03 by a single checkbox change in any statistical software. The mathematical operation is trivial. The inferential validity depends entirely on whether the directional hypothesis was declared before the data were analyzed. Without a record of when the hypothesis was declared, there is no way to distinguish valid one-sided testing from retrospective selection of the favorable tail.

The practical consequence: one-sided tests reported without rationale are increasingly treated by reviewers as a red flag. Several journals (including BMJ and JAMA) now discourage one-sided testing except in specific pre-specified contexts (non-inferiority trials, superiority margins).

[Image placeholder: Side-by-side t-distribution curves showing two-sided test (alpha split evenly at both tails, critical region at ﾂｱ1.96) vs. one-sided test (full alpha at one tail, critical value at 1.645). The same observed statistic of 1.75 is shown —non-significant two-sided (p=0.081), significant one-sided (p=0.040).]

The same observed statistic yields two different inferential conclusions depending on test directionality. Without a pre-declared hypothesis, the choice of directionality is an analytical decision that inflates false positive risk.

STEP 2 —Journal Requirement

CONSORT requires one-sided tests to be pre-specified with rationale in clinical trial registrations. ICMJE guidelines discourage one-sided tests except when justified by prior evidence or clear mechanistic reasoning. The BMJ requires explicit justification in a covering letter when one-sided tests are used. Most journals require one-sided hypothesis and direction to appear in the pre-registration or Methods section, not just in results.

STEP 3 —Licklider Solution

Input

Test directionality selection: two-sided (default) or one-sided (requires additional input)
For one-sided: declared direction (A > B or A < B), source of prior hypothesis (citation, mechanistic reasoning, or pre-registration ID)
Timestamp of declaration (recorded automatically)

Output

Two-sided p-value always calculated and displayed alongside any one-sided result
One-sided p-value displayed with explicit direction label: "p (one-sided, A > B) = 0.031"
Rationale recorded in the analysis log and included in auto-generated Methods text
If one-sided result is significant but two-sided is not: a disclosure flag is auto-generated and included in export

Guard

One-sided p-values cannot be calculated until the directional hypothesis and rationale are recorded. If the user selects "one-sided" without providing a rationale, the system displays the rationale input dialog and does not proceed. For rank-based one-sided tests, treat product behavior as not verified end-to-end — see Non-Parametric Alternatives. Exploratory analyses marked via Outcome Type Lock are not blocked but receive an advisory note.

[Image placeholder: Licklider one-sided test dialog showing: direction selection (Group A > Group B / Group A < Group B), hypothesis source field (text input: "Based on mechanistic evidence from Smith et al. 2023 showing X increases Y"), and a note that the two-sided p-value will be displayed in parallel. The "Calculate one-sided p" button is grayed out until both fields are complete.]

One-sided test lock: direction and rationale required before one-sided p-value is calculated. Two-sided p-value always shown in parallel.

STEP 4 —Draft Output (Draft / Needs review)

A one-sided t-test was used based on an a priori directional hypothesis
(Group A > Group B) derived from mechanistic evidence in prior work
(Smith et al., 2023, Nature Cell Biology). The one-sided p-value is
reported; the two-sided p-value is provided for reference
(p_one-sided = 0.031; p_two-sided = 0.062). The directional hypothesis
and test configuration were pre-declared prior to data analysis.