Start at the certainty level implied by the study designs, then apply the five downgrading domains and record a reason for each judgement at the outcome level. GRADE (Grading of Recommendations Assessment, Development and Evaluation) rates certainty per outcome, so a winter-health review can give different ratings to, for example, hospital admissions and symptom duration.
Set the starting level
Randomized evidence begins as high certainty and observational evidence begins as low certainty. Record which design supports the outcome before you judge anything else.
Apply the five downgrading domains
- Assess risk of bias across the studies contributing to the outcome; downgrade when limitations in randomization, blinding, or attrition could change the result, and note which studies drove the concern.
- Assess inconsistency by comparing effect direction and size across studies; downgrade when the results conflict and no subgroup explains the difference.
- Assess indirectness by checking whether the populations, interventions, and outcomes match your review question; downgrade when the evidence answers a related but different question.
- Assess imprecision by examining confidence intervals and event counts; downgrade when the estimate is too wide to support a decision.
- Assess publication bias using funnel plots or small-study effects where enough studies exist; downgrade when the pattern suggests missing results.
Record the trail
For each domain, write the judgement, the reason, and the studies involved. The final rating is the starting level minus the domains you downgraded, reported per outcome. A footnote trail that names the domain, the reason, and the supporting studies lets another reviewer audit the rating and reproduce it. Cochrane's GRADE resources and Summary of Findings guidance describe this outcome-level structure, and GRADEpro supports building the table.