study-design

name	study-design
description	Convert a qualified research brief and evidence ledger into a testable study-design contract with an explicit estimand, measurement plan, bias analysis, analysis plan, falsifiers, and governance gates. Use after topic qualification and before data analysis or manuscript drafting.

Study Design

Design the smallest credible study that answers the qualified question. Do not add methods because they are fashionable or conventional.

Required inputs

Require:

• research-brief.md with a bounded CONTINUE or specified PIVOT;
• evidence-ledger.json with the minimum canonical fields;
• verified information about available data or planned collection.

Use ../../templates/study-design.md. If a required input is absent, invalid, or blocked, set the design status to BLOCKED and identify the exact repair action.

Evidence and inference boundary

Anchor design choices to the research brief, evidence ledger, data documentation, or user decisions. Mark unsupported assumptions explicitly. Never invent data availability, sample size, event rates, measurement properties, approvals, registrations, or power.

Design workflow

1. Express the question as a population, focal condition or exposure, comparator, outcome, and time horizon.
2. State the target estimand or descriptive quantity and its interpretation.
3. Choose the design and explain why it identifies or describes that target.
4. Lock time zero, eligibility, follow-up, measurement timing, and unit of analysis.
5. Classify variables as exposure, outcome, confounder, effect modifier, precision variable, mediator, or excluded variable. Do not adjust for a mediator or collider without an explicit rationale.
6. Enumerate identification assumptions and threats: confounding, selection, attrition, measurement error, missingness, multiplicity, and leakage.
7. Define the primary analysis, uncertainty estimate, diagnostics, sensitivity analyses, negative controls, and falsification tests.
8. Build a risk register that names each threat or likely review attack, observable trigger, impact, contingency, required evidence, and owner or human decision. Do not describe a contingency as completed work.
9. Define reproducibility, protocol-deviation, ethics, privacy, and data-use gates.

Decision

Choose:

• DESIGN_READY: all non-compensable gates pass;
• REVISE: a specific design change can close the remaining gap;
• BLOCKED: data, measurement, identification, ethics, or feasibility cannot currently support the question.

Any parser or tool failure is evidence that a check was not completed, not a pass. Do not silently fill or auto-repair missing inputs.

Call record_checkpoint with the decisive assumption, uncertainty, or stop rule. Prefix the statement with [study-design].

Output

Write study-design.md from the template and return a concise summary of:

1. question and estimand;
2. design and measurement;
3. bias and identification assumptions;
4. primary and sensitivity analyses;
5. risk register and contingencies;
6. reproducibility and governance;
7. DESIGN_READY, REVISE, or BLOCKED;
8. unresolved gates and evidence needed.