/mstack:hypothesis-design

Stage: map · Voice: methodologist

After /mstack:theory-build; before fielding, analysis, or /mstack:preregister (which quotes these hypotheses verbatim). The bar is the one the preregistration will be held to: two researchers given the hypothesis and the data should test the same thing (see ${CLAUDE_PLUGIN_ROOT}/references/preregistration-guide.md, section "What must be specific").

Procedure

1. Load .mstack/theory.md (mechanism, rivals, estimand) and .mstack/lit-map.md (comparable effect sizes).

2. Specify eight fields per hypothesis (H1, H2, …):

   Field	Example
   Statement	"Higher exposure to import competition increases vote share for protectionist parties."
   Direction	Positive (β > 0).
   Estimand	"The ATT of a one-SD increase in exposure on protectionist vote share (percentage points), among NUTS-2 regions in 12 EU states, 2000–2019" (Lundberg, Johnson & Stewart 2021).
   Magnitude expectation	"A 1-SD increase in exposure raises vote share by 1–3 percentage points; comparable estimates in Colantone & Stanig (2018) are …"
   Smallest effect of substantive interest	"1 percentage point; smaller effects would not change the policy or theoretical claim."
   Operationalization of X	"ADH-style instrument as in Autor et al. 2013, country-region-year level."
   Operationalization of Y	"Percent of constituency vote for parties CMP classifies as protectionist."
   Falsification pattern and null reading	"A 95% CI excluding zero supports H1; a CI entirely within (−1, 1) is evidence against a substantively meaningful effect; a CI straddling both is inconclusive. β < 0 falsifies the mechanism direction."

3. Falsifiability check. Could a result actually falsify it, or is the prediction "some relationship"? Is the falsifying pattern observable with the planned data at the planned N (check against .mstack/power-analysis.md once it exists)? If not, the hypothesis is untestable in this design.
4. Exactly one primary. The headline rests on it; secondaries explore mechanism, heterogeneity, or scope. Every secondary carries the same eight fields.
5. Moderation and mechanism. A theorized moderator is its own hypothesis stating the marginal effect of X at each level of the moderator, not the interaction coefficient ("H2: the effect of X on Y is at least twice as large where S is high as where S is low"); a theorized mediator is a hypothesis about the path, with the caveat that mediation analysis needs its own identifying assumptions. Heterogeneity not specified here is exploratory in the prereg. Take the rows of the rival-mechanisms table in theory.md and write, for each, the hypothesis that would distinguish your mechanism from that rival.
6. Save to .mstack/hypotheses.md; the prereg pulls from it.

Outputs

• .mstack/hypotheses.md — every hypothesis with all eight fields, the primary marked, and the rival-discriminating hypotheses grouped at the end.
• Summary block: primary hypothesis, its estimand, its SESOI, and its falsification pattern.

Anti-patterns

• One-sided ambiguity. "X is related to Y" is not a hypothesis.
• No estimand. "The coefficient on X" is not a quantity about the world.
• Hypotheses bigger than the data. If the data cannot show the falsifying pattern, get different data or rewrite the hypothesis.
• Five primaries. Exactly one.
• A null with no interpretation. Decide now what a null result will be taken to mean.

Next

/mstack:identification-review, then /mstack:preregister.