From question to analysis plan

You will learn to choose the health question a table can answer, then write a short plan that makes the answer checkable. The outcome is not a test name: it is a stated comparison, an estimated effect, and an honest account of uncertainty.

Start with the clinic decision

A clinic adds six weeks of discharge support. Its follow-up table has one row per person: group, symptom score at discharge, and symptom score six weeks later. The team must decide what to report before reading a result.

• Question A: At six weeks, how do the support and usual-care groups differ?
• Question B: How much did each person’s symptoms change from discharge?
• Question C: How do the groups differ in their change from discharge?

Choose Question A when the clinic’s decision is which group had lower symptoms at the scheduled follow-up. Question B describes change within people; it does not compare the groups. Question C compares change, so it is appropriate only when that is the decision the clinic wants to make. The same table can support all three, but it cannot make them the same question.

How the same question may appear in a paper

You have already stated the population, comparison, outcome, and time in ordinary language. Some papers compress those parts into PICO: Population, Intervention, Comparator, Outcome. Here that means eligible adults leaving the clinic; the support program; usual care; and symptom score at six weeks. For a non-randomized exposure question, PECO changes Intervention to Exposure. Use the acronym as a reading aid, not as a substitute for the complete question.

The estimand is the exact effect or question the data must estimate. For Question A: the mean symptom-score difference at six weeks, support minus usual care, among eligible adults. An estimate is the number calculated from this sample for that target. The unit of analysis is what one row contributes: here, one participant, not one clinic visit. A confidence interval (CI) is the range of effect sizes compatible with the data and stated assumptions; it shows uncertainty, not whether the program is clinically proven.

Write a plain-language analysis plan

A plan records the choices before a result can tempt you to change the question. Begin with ordinary language:

The plan chooses Question A. “Before and after” would answer Question B instead; “change in each group” would answer Question C unless it explicitly compares those changes. A negative support-minus-usual-care estimate would mean lower average symptoms in the support group; the CI would show how uncertain that estimated difference remains. This is a teaching decision, not evidence that the program works.

The canonical 12-item method audit

This numbered checklist is the stable audit used throughout the track. Answer each item in plain language before treating a method or result as ready; a missing or conflicting answer requires review.

1. Target effect (estimand). State the exact contrast or quantity the analysis is meant to estimate.
2. Population and unit. Name who can contribute and what one independent observation represents.
3. Outcome scale. Define what is measured, its unit or range, and what higher and lower values mean.
4. Groups and times. Name the groups, conditions, and time points that form the comparison.
5. Dependence. Identify repeated measurements, shared participants, sites, families, or clusters that make observations related.
6. Sampling and bias. Explain how records enter the analysis and how selection, exclusions, or missingness could distort the target effect.
7. Data defects. Check impossible values, duplicates, units, dates, and cross-field inconsistencies while retaining an audit trail.
8. Assumptions. State the model and missing-data assumptions and how their consequences will be challenged.
9. Pre-specification and multiplicity. Name the primary claim before results and how the full family of comparisons will be handled.
10. Direct-effect method. Choose a method that directly estimates item 1 rather than a convenient proxy.
11. Effect, interval, and diagnostics. Report magnitude, uncertainty, and relevant model or data checks together.
12. Causal and general limits. State what the design cannot establish about cause and about populations beyond the observed sample.

The item number is part of the contract: later modules link back here instead of silently renumbering the audit.

Try a new decision without code

A community clinic must choose between phone and text-message check-ins for people starting blood-pressure medicine. Its table has one row per person: check-in format and whether the person is taking the medicine as prescribed 14 days later. Write five lines: population, comparison, outcome and time, unit, and effect.

Keep later complications in view, not at the entrance

Later modules add data integrity, missing-data assumptions, diagnosis, and reporting. In trials, an intercurrent event—for example, stopping treatment or starting rescue therapy—can change which question is being answered. ICH E9(R1) is a public extension for that planning problem. For now, record such events and do not silently switch the question after results appear.

Optional technical practice: AURORA-30

AURORA-30 is a fictional practice dossier made from small synthetic fixtures. It contains no real participants, institutions, interventions, or clinical outcomes, and results from it are not clinical evidence. You do not need it, JSON, or code to understand this module. A count or value carries forward only when the next module names the same file and field; the clinic example at six weeks above is a separate no-code example.

The optional lab below checks the day-30 planning branch: whether a synthetic plan names its target, unit, time, and evidence.

{
  "population": "eligible synthetic participants",
  "comparison": "support minus usual care",
  "outcome": "symptom score at day 30",
  "unit": "participant",
  "estimand": "mean difference at day 30"
}