From question to analysis plan

“Did the program help?” is not one statistical question. AURORA-30 is the wholly synthetic teaching cohort and dataset used to make that conflict visible: 30 fictional records, with no real patients, institutions, interventions, or outcomes; it does not constitute clinical evidence.

The conflict: one table, three defensible routes

The same AURORA-30 columns support three plausible analyses:

• Day 30 between groups: estimate the group difference at the end of follow-up.
• Before and after in the same person: estimate within-person change from baseline.
• Change between groups: estimate the difference between the groups’ mean changes.

Each route can be mathematically correct, yet each answers a different question. The contrast, dependence structure, denominator, and interpretation change. A test name cannot decide which answer the study owes its reader.

The move: make the decisions recoverable

An analysis contract is not a legal document or an approval form. It is versioned, verifiable memory of the decisions that make the result mean what it claims: population, contrast, outcome, time, unit, effect measure, missing-data rule, and intended interpretation. It lets another analyst reconstruct the route, challenge a choice, and distinguish a documented revision from a silent rewrite.

PICO or PECO frames the question; the estimand closes it

PICO records population, intervention, comparator, and outcome. For observational questions, PECO replaces intervention with exposure. Add the time horizon, unit of analysis, effect measure, and handling of intercurrent events—post-baseline occurrences, such as treatment discontinuation or rescue treatment, that affect how the outcome should be interpreted or whether it can be observed. That last move turns a topic into a computation-ready contract.

For the first route, the teaching question becomes: among eligible synthetic participants, what is the 30-day difference in mean symptom score between program and usual-care groups, analyzed at participant level under the pre-specified missing-data rule? The effect is a mean difference with a confidence interval. The other two routes remain legitimate questions, but they are not this estimand.

Worked example: write the AURORA-30 contract

analysis_plan = {
  "population": "eligible AURORA-30 participants",
  "unit": "participant",
  "exposure": "discharge_program",
  "comparison": "usual_care",
  "outcome": "symptom_score_day_30",
  "estimand": "mean_difference",
  "time_horizon_days": 30,
  "analysis_set": "all_assigned_with_prespecified_missing_rule",
  "evidence_bundle": ["effect", "confidence_interval", "diagnostic", "limitation"]
}

assert analysis_plan["unit"] == "participant"
assert "effect" in analysis_plan["evidence_bundle"]

The contract deliberately does not name a test. The design and estimand constrain the candidate methods; HEALTH.2–4 determine whether the data can support them. In HEALTH.5 the hypothesis becomes a secondary statement about the estimand, not the owner of the analysis.

Your turn: freeze the question before opening the data

Retrieve the primer's design-before-method decision: the question and dependence structure must constrain the method, never the other way around. Write one sentence that contains population, comparison, outcome, time, unit, and effect measure. Then answer the first six items of the universal method checklist: estimand, population/unit, outcome scale, groups/times, dependence, and sampling bias. If any answer is “unknown,” the plan requires review (HOLD)—not because the researcher or study failed, but because the question is not yet executable.

Decision audit: can another analyst reproduce the question?

A useful plan resolves predictable forks: analysis by assignment or observed exposure, reduction of repeated records to the participant unit, acceptable day-30 window, and the primary rule when outcomes are missing. It also separates descriptive, associational, and causal intent. PICO or PECO organizes a question; it does not eliminate selection bias, create exchangeability, or guarantee generalization. Put the intended interpretation beside the estimand so later modules cannot inflate it.

Run a handoff test. Give the contract—without the dataset—to another analyst. They should select the same population, row unit, outcome, time, contrast direction, and effect measure. If two defensible analyses remain possible, the contract is not finished. Record those forks before proceeding rather than allowing software defaults to decide them.

Version the plan. A revision made before outcome inspection differs from a change triggered by the observed effect. Preserve old and new versions, the reason, who approved it, and whether the estimand changed. HEALTH.12 will compare the final report with the version that actually governed the work. Pre-specification is not rigidity; it is visible accountability for deviations.

Name the contrast direction explicitly: program minus usual care is not interchangeable with its inverse.

Label every AURORA-30 quantity as synthetic and educational so no reader mistakes the exercise for clinical evidence.