Causal Paper to Advice
Turn a paper into an advice chain:
intervention → comparison → outcome → effect → boundaries → recommendation
Every link must be supported. A broken link weakens everything to its right.
Establish the claim
Read the paper itself, including the methods, tables or figures that support the main claim and the limitations. Use commentary, press releases and coverage only to locate or challenge the paper, never as substitutes for it.
State the claim in one sentence:
For population, changing exposure/intervention from counterfactual changes outcome over time horizon by effect estimate.
If any slot is missing, say so. Record whether the paper is peer-reviewed, a working paper, preprint, correspondence or other publication type; this affects confidence, not the result by itself.
Reconstruct identification
Explain why the comparison could reveal causality in plain language. Name the design only after describing the assignment or quasi-random variation that does the identifying work.
Trace:
- who received the intervention or exposure;
- who serves as the counterfactual;
- why the groups would otherwise have followed comparable paths;
- what outcome was measured, when and from which data;
- which assumptions must hold for the estimate to be causal.
Then attack the comparison with the strongest plausible rival explanation: selection, confounding, reverse causality, differential measurement, attrition, spillovers, anticipation, policy changes occurring at the same time or selective reporting. Report the authors’ test or design feature that addresses it and what remains unresolved.
Treat dose-response, timing gradients, sibling comparisons, negative controls, placebo tests and robustness checks as supporting evidence. None independently manufactures random assignment.
When the design cannot identify causality, relabel the finding as an association and continue only if an associative result is still useful.
Size the effect
Give the estimate in the units a person or decision-maker experiences. Preserve the statistical estimate, then translate it.
Include, where available:
- baseline risk or outcome level;
- absolute change and relative change;
- time horizon;
- uncertainty interval, not merely statistical significance;
- average effect and meaningful subgroup effects;
- take-up, adherence or treatment intensity;
- harms, costs and displaced outcomes.
Use a common denominator for risks. For example, convert “20% lower risk” into events per 1,000 when baseline risk is available. Identify extrapolations explicitly, including estimates projected from a per-year effect to a whole childhood or from trial efficacy to population impact.
Distinguish “detectable” from “important”. Compare the magnitude with a relevant benchmark: baseline risk, current practice, another feasible intervention, cost or burden of action.
Map the boundaries
Build an edge map before recommending anything:
- People: age, sex, baseline risk, disease severity, geography, income or other relevant subgroup.
- Treatment: dose, duration, delivery, co-interventions and adherence.
- Outcome: surrogate versus lived outcome; intended benefits versus spillovers or harms.
- Time: immediate versus delayed effects; persistence after treatment.
- Setting: institutions, incentives and resources required for the result to travel.
Actively look for reversals and nulls. A benefit for younger children does not imply a benefit for teenagers; a full medical dose does not establish the value of a microdose; a programme that helps participants may still harm people it displaces.
Separate three questions:
- Did the intervention cause the outcome?
- Which component or mechanism caused it?
- Would a feasible real-world version preserve that component without importing unacceptable harms?
Evidence for the first does not automatically answer the next two.
Derive advice
Choose the strongest recommendation licensed by the advice chain:
- Act: benefits, harms and practical conditions are sufficiently established for the target audience.
- Consider conditionally: evidence is useful for a defined group or under supervision, monitoring or resource constraints.
- Do not infer: the result is associative, the tested intervention differs materially from the proposed action, or important harms are unmeasured.
- Research first: the key comparison, dose, subgroup or implementation question has not been tested.
State the audience, action, conditions and confidence. Prefer the smallest action that preserves the demonstrated mechanism. When the whole programme has serious unmeasured costs, recommend the transferable ingredient rather than the programme itself.
For health advice, distinguish population guidance from individual medical decisions. State clinically important contraindications or supervision requirements reported by the evidence. Do not convert a biomarker association into a treatment claim or a population average into a personal diagnosis.
Deliverable
Lead with the bottom line, then provide:
- What the paper means — the completed causal sentence.
- Why it may be causal — comparison, identifying variation and assumptions.
- How large the effect is — absolute and relative magnitude, uncertainty and benchmark.
- Where it breaks — the edge map, rival explanations and adverse outcomes.
- What follows — recommendation tier, audience, conditions and confidence.
- What would change the conclusion — the most decision-relevant missing evidence.
Cite the paper precisely and attach each important number to its table, figure or page when possible. Clearly label calculations and interpretations made during the analysis.
The analysis is complete when every link in the advice chain is either supported or visibly marked missing, the main estimate is expressed at human scale, the strongest rival explanation and consequential edge case have been tested, and the recommendation does not outrun the weakest link.